1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Extent allocs and frees
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
31 #include <linux/quotaops.h>
32 #include <linux/blkdev.h>
34 #include <cluster/masklog.h>
40 #include "blockcheck.h"
42 #include "extent_map.h"
45 #include "localalloc.h"
52 #include "refcounttree.h"
53 #include "ocfs2_trace.h"
55 #include "buffer_head_io.h"
57 enum ocfs2_contig_type {
64 static enum ocfs2_contig_type
65 ocfs2_extent_rec_contig(struct super_block *sb,
66 struct ocfs2_extent_rec *ext,
67 struct ocfs2_extent_rec *insert_rec);
69 * Operations for a specific extent tree type.
71 * To implement an on-disk btree (extent tree) type in ocfs2, add
72 * an ocfs2_extent_tree_operations structure and the matching
73 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it
74 * for the allocation portion of the extent tree.
76 struct ocfs2_extent_tree_operations {
78 * last_eb_blk is the block number of the right most leaf extent
79 * block. Most on-disk structures containing an extent tree store
80 * this value for fast access. The ->eo_set_last_eb_blk() and
81 * ->eo_get_last_eb_blk() operations access this value. They are
84 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
86 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
89 * The on-disk structure usually keeps track of how many total
90 * clusters are stored in this extent tree. This function updates
91 * that value. new_clusters is the delta, and must be
92 * added to the total. Required.
94 void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
98 * If this extent tree is supported by an extent map, insert
99 * a record into the map.
101 void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
102 struct ocfs2_extent_rec *rec);
105 * If this extent tree is supported by an extent map, truncate the
108 void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
112 * If ->eo_insert_check() exists, it is called before rec is
113 * inserted into the extent tree. It is optional.
115 int (*eo_insert_check)(struct ocfs2_extent_tree *et,
116 struct ocfs2_extent_rec *rec);
117 int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
120 * --------------------------------------------------------------
121 * The remaining are internal to ocfs2_extent_tree and don't have
126 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
129 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
132 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
133 * it exists. If it does not, et->et_max_leaf_clusters is set
134 * to 0 (unlimited). Optional.
136 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
139 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
140 * are contiguous or not. Optional. Don't need to set it if use
141 * ocfs2_extent_rec as the tree leaf.
143 enum ocfs2_contig_type
144 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
145 struct ocfs2_extent_rec *ext,
146 struct ocfs2_extent_rec *insert_rec);
151 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
154 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
155 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
157 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
159 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
160 struct ocfs2_extent_rec *rec);
161 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
163 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
164 struct ocfs2_extent_rec *rec);
165 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
166 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
167 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
168 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk,
169 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk,
170 .eo_update_clusters = ocfs2_dinode_update_clusters,
171 .eo_extent_map_insert = ocfs2_dinode_extent_map_insert,
172 .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
173 .eo_insert_check = ocfs2_dinode_insert_check,
174 .eo_sanity_check = ocfs2_dinode_sanity_check,
175 .eo_fill_root_el = ocfs2_dinode_fill_root_el,
178 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
181 struct ocfs2_dinode *di = et->et_object;
183 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
184 di->i_last_eb_blk = cpu_to_le64(blkno);
187 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
189 struct ocfs2_dinode *di = et->et_object;
191 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
192 return le64_to_cpu(di->i_last_eb_blk);
195 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
198 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
199 struct ocfs2_dinode *di = et->et_object;
201 le32_add_cpu(&di->i_clusters, clusters);
202 spin_lock(&oi->ip_lock);
203 oi->ip_clusters = le32_to_cpu(di->i_clusters);
204 spin_unlock(&oi->ip_lock);
207 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
208 struct ocfs2_extent_rec *rec)
210 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
212 ocfs2_extent_map_insert_rec(inode, rec);
215 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
218 struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
220 ocfs2_extent_map_trunc(inode, clusters);
223 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
224 struct ocfs2_extent_rec *rec)
226 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
227 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
229 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
230 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
231 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
232 "Device %s, asking for sparse allocation: inode %llu, "
233 "cpos %u, clusters %u\n",
235 (unsigned long long)oi->ip_blkno,
236 rec->e_cpos, oi->ip_clusters);
241 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
243 struct ocfs2_dinode *di = et->et_object;
245 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
246 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
251 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
253 struct ocfs2_dinode *di = et->et_object;
255 et->et_root_el = &di->id2.i_list;
259 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
261 struct ocfs2_xattr_value_buf *vb = et->et_object;
263 et->et_root_el = &vb->vb_xv->xr_list;
266 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
269 struct ocfs2_xattr_value_buf *vb = et->et_object;
271 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
274 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
276 struct ocfs2_xattr_value_buf *vb = et->et_object;
278 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
281 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
284 struct ocfs2_xattr_value_buf *vb = et->et_object;
286 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
289 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
290 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk,
291 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk,
292 .eo_update_clusters = ocfs2_xattr_value_update_clusters,
293 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el,
296 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
298 struct ocfs2_xattr_block *xb = et->et_object;
300 et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
303 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
305 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
306 et->et_max_leaf_clusters =
307 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
310 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
313 struct ocfs2_xattr_block *xb = et->et_object;
314 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
316 xt->xt_last_eb_blk = cpu_to_le64(blkno);
319 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
321 struct ocfs2_xattr_block *xb = et->et_object;
322 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
324 return le64_to_cpu(xt->xt_last_eb_blk);
327 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
330 struct ocfs2_xattr_block *xb = et->et_object;
332 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
335 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
336 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk,
337 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk,
338 .eo_update_clusters = ocfs2_xattr_tree_update_clusters,
339 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el,
340 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
343 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
346 struct ocfs2_dx_root_block *dx_root = et->et_object;
348 dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
351 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
353 struct ocfs2_dx_root_block *dx_root = et->et_object;
355 return le64_to_cpu(dx_root->dr_last_eb_blk);
358 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
361 struct ocfs2_dx_root_block *dx_root = et->et_object;
363 le32_add_cpu(&dx_root->dr_clusters, clusters);
366 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
368 struct ocfs2_dx_root_block *dx_root = et->et_object;
370 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
375 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
377 struct ocfs2_dx_root_block *dx_root = et->et_object;
379 et->et_root_el = &dx_root->dr_list;
382 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
383 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk,
384 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk,
385 .eo_update_clusters = ocfs2_dx_root_update_clusters,
386 .eo_sanity_check = ocfs2_dx_root_sanity_check,
387 .eo_fill_root_el = ocfs2_dx_root_fill_root_el,
390 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
392 struct ocfs2_refcount_block *rb = et->et_object;
394 et->et_root_el = &rb->rf_list;
397 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
400 struct ocfs2_refcount_block *rb = et->et_object;
402 rb->rf_last_eb_blk = cpu_to_le64(blkno);
405 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
407 struct ocfs2_refcount_block *rb = et->et_object;
409 return le64_to_cpu(rb->rf_last_eb_blk);
412 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
415 struct ocfs2_refcount_block *rb = et->et_object;
417 le32_add_cpu(&rb->rf_clusters, clusters);
420 static enum ocfs2_contig_type
421 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
422 struct ocfs2_extent_rec *ext,
423 struct ocfs2_extent_rec *insert_rec)
428 static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
429 .eo_set_last_eb_blk = ocfs2_refcount_tree_set_last_eb_blk,
430 .eo_get_last_eb_blk = ocfs2_refcount_tree_get_last_eb_blk,
431 .eo_update_clusters = ocfs2_refcount_tree_update_clusters,
432 .eo_fill_root_el = ocfs2_refcount_tree_fill_root_el,
433 .eo_extent_contig = ocfs2_refcount_tree_extent_contig,
436 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
437 struct ocfs2_caching_info *ci,
438 struct buffer_head *bh,
439 ocfs2_journal_access_func access,
441 struct ocfs2_extent_tree_operations *ops)
446 et->et_root_journal_access = access;
448 obj = (void *)bh->b_data;
451 et->et_ops->eo_fill_root_el(et);
452 if (!et->et_ops->eo_fill_max_leaf_clusters)
453 et->et_max_leaf_clusters = 0;
455 et->et_ops->eo_fill_max_leaf_clusters(et);
458 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
459 struct ocfs2_caching_info *ci,
460 struct buffer_head *bh)
462 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
463 NULL, &ocfs2_dinode_et_ops);
466 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
467 struct ocfs2_caching_info *ci,
468 struct buffer_head *bh)
470 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
471 NULL, &ocfs2_xattr_tree_et_ops);
474 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
475 struct ocfs2_caching_info *ci,
476 struct ocfs2_xattr_value_buf *vb)
478 __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
479 &ocfs2_xattr_value_et_ops);
482 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
483 struct ocfs2_caching_info *ci,
484 struct buffer_head *bh)
486 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
487 NULL, &ocfs2_dx_root_et_ops);
490 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
491 struct ocfs2_caching_info *ci,
492 struct buffer_head *bh)
494 __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
495 NULL, &ocfs2_refcount_tree_et_ops);
498 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
501 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
504 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
506 return et->et_ops->eo_get_last_eb_blk(et);
509 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
512 et->et_ops->eo_update_clusters(et, clusters);
515 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
516 struct ocfs2_extent_rec *rec)
518 if (et->et_ops->eo_extent_map_insert)
519 et->et_ops->eo_extent_map_insert(et, rec);
522 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
525 if (et->et_ops->eo_extent_map_truncate)
526 et->et_ops->eo_extent_map_truncate(et, clusters);
529 static inline int ocfs2_et_root_journal_access(handle_t *handle,
530 struct ocfs2_extent_tree *et,
533 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
537 static inline enum ocfs2_contig_type
538 ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
539 struct ocfs2_extent_rec *rec,
540 struct ocfs2_extent_rec *insert_rec)
542 if (et->et_ops->eo_extent_contig)
543 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
545 return ocfs2_extent_rec_contig(
546 ocfs2_metadata_cache_get_super(et->et_ci),
550 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
551 struct ocfs2_extent_rec *rec)
555 if (et->et_ops->eo_insert_check)
556 ret = et->et_ops->eo_insert_check(et, rec);
560 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
564 if (et->et_ops->eo_sanity_check)
565 ret = et->et_ops->eo_sanity_check(et);
569 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
570 struct ocfs2_extent_block *eb);
571 static void ocfs2_adjust_rightmost_records(handle_t *handle,
572 struct ocfs2_extent_tree *et,
573 struct ocfs2_path *path,
574 struct ocfs2_extent_rec *insert_rec);
576 * Reset the actual path elements so that we can re-use the structure
577 * to build another path. Generally, this involves freeing the buffer
580 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
582 int i, start = 0, depth = 0;
583 struct ocfs2_path_item *node;
588 for(i = start; i < path_num_items(path); i++) {
589 node = &path->p_node[i];
597 * Tree depth may change during truncate, or insert. If we're
598 * keeping the root extent list, then make sure that our path
599 * structure reflects the proper depth.
602 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
604 path_root_access(path) = NULL;
606 path->p_tree_depth = depth;
609 void ocfs2_free_path(struct ocfs2_path *path)
612 ocfs2_reinit_path(path, 0);
618 * All the elements of src into dest. After this call, src could be freed
619 * without affecting dest.
621 * Both paths should have the same root. Any non-root elements of dest
624 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
628 BUG_ON(path_root_bh(dest) != path_root_bh(src));
629 BUG_ON(path_root_el(dest) != path_root_el(src));
630 BUG_ON(path_root_access(dest) != path_root_access(src));
632 ocfs2_reinit_path(dest, 1);
634 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
635 dest->p_node[i].bh = src->p_node[i].bh;
636 dest->p_node[i].el = src->p_node[i].el;
638 if (dest->p_node[i].bh)
639 get_bh(dest->p_node[i].bh);
644 * Make the *dest path the same as src and re-initialize src path to
647 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
651 BUG_ON(path_root_bh(dest) != path_root_bh(src));
652 BUG_ON(path_root_access(dest) != path_root_access(src));
654 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
655 brelse(dest->p_node[i].bh);
657 dest->p_node[i].bh = src->p_node[i].bh;
658 dest->p_node[i].el = src->p_node[i].el;
660 src->p_node[i].bh = NULL;
661 src->p_node[i].el = NULL;
666 * Insert an extent block at given index.
668 * This will not take an additional reference on eb_bh.
670 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
671 struct buffer_head *eb_bh)
673 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
676 * Right now, no root bh is an extent block, so this helps
677 * catch code errors with dinode trees. The assertion can be
678 * safely removed if we ever need to insert extent block
679 * structures at the root.
683 path->p_node[index].bh = eb_bh;
684 path->p_node[index].el = &eb->h_list;
687 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
688 struct ocfs2_extent_list *root_el,
689 ocfs2_journal_access_func access)
691 struct ocfs2_path *path;
693 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
695 path = kzalloc(sizeof(*path), GFP_NOFS);
697 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
699 path_root_bh(path) = root_bh;
700 path_root_el(path) = root_el;
701 path_root_access(path) = access;
707 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
709 return ocfs2_new_path(path_root_bh(path), path_root_el(path),
710 path_root_access(path));
713 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
715 return ocfs2_new_path(et->et_root_bh, et->et_root_el,
716 et->et_root_journal_access);
720 * Journal the buffer at depth idx. All idx>0 are extent_blocks,
721 * otherwise it's the root_access function.
723 * I don't like the way this function's name looks next to
724 * ocfs2_journal_access_path(), but I don't have a better one.
726 int ocfs2_path_bh_journal_access(handle_t *handle,
727 struct ocfs2_caching_info *ci,
728 struct ocfs2_path *path,
731 ocfs2_journal_access_func access = path_root_access(path);
734 access = ocfs2_journal_access;
737 access = ocfs2_journal_access_eb;
739 return access(handle, ci, path->p_node[idx].bh,
740 OCFS2_JOURNAL_ACCESS_WRITE);
744 * Convenience function to journal all components in a path.
746 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
748 struct ocfs2_path *path)
755 for(i = 0; i < path_num_items(path); i++) {
756 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
768 * Return the index of the extent record which contains cluster #v_cluster.
769 * -1 is returned if it was not found.
771 * Should work fine on interior and exterior nodes.
773 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
777 struct ocfs2_extent_rec *rec;
778 u32 rec_end, rec_start, clusters;
780 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
781 rec = &el->l_recs[i];
783 rec_start = le32_to_cpu(rec->e_cpos);
784 clusters = ocfs2_rec_clusters(el, rec);
786 rec_end = rec_start + clusters;
788 if (v_cluster >= rec_start && v_cluster < rec_end) {
798 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
799 * ocfs2_extent_rec_contig only work properly against leaf nodes!
801 static int ocfs2_block_extent_contig(struct super_block *sb,
802 struct ocfs2_extent_rec *ext,
805 u64 blk_end = le64_to_cpu(ext->e_blkno);
807 blk_end += ocfs2_clusters_to_blocks(sb,
808 le16_to_cpu(ext->e_leaf_clusters));
810 return blkno == blk_end;
813 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
814 struct ocfs2_extent_rec *right)
818 left_range = le32_to_cpu(left->e_cpos) +
819 le16_to_cpu(left->e_leaf_clusters);
821 return (left_range == le32_to_cpu(right->e_cpos));
824 static enum ocfs2_contig_type
825 ocfs2_extent_rec_contig(struct super_block *sb,
826 struct ocfs2_extent_rec *ext,
827 struct ocfs2_extent_rec *insert_rec)
829 u64 blkno = le64_to_cpu(insert_rec->e_blkno);
832 * Refuse to coalesce extent records with different flag
833 * fields - we don't want to mix unwritten extents with user
836 if (ext->e_flags != insert_rec->e_flags)
839 if (ocfs2_extents_adjacent(ext, insert_rec) &&
840 ocfs2_block_extent_contig(sb, ext, blkno))
843 blkno = le64_to_cpu(ext->e_blkno);
844 if (ocfs2_extents_adjacent(insert_rec, ext) &&
845 ocfs2_block_extent_contig(sb, insert_rec, blkno))
852 * NOTE: We can have pretty much any combination of contiguousness and
855 * The usefulness of APPEND_TAIL is more in that it lets us know that
856 * we'll have to update the path to that leaf.
858 enum ocfs2_append_type {
863 enum ocfs2_split_type {
869 struct ocfs2_insert_type {
870 enum ocfs2_split_type ins_split;
871 enum ocfs2_append_type ins_appending;
872 enum ocfs2_contig_type ins_contig;
873 int ins_contig_index;
877 struct ocfs2_merge_ctxt {
878 enum ocfs2_contig_type c_contig_type;
879 int c_has_empty_extent;
880 int c_split_covers_rec;
883 static int ocfs2_validate_extent_block(struct super_block *sb,
884 struct buffer_head *bh)
887 struct ocfs2_extent_block *eb =
888 (struct ocfs2_extent_block *)bh->b_data;
890 trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
892 BUG_ON(!buffer_uptodate(bh));
895 * If the ecc fails, we return the error but otherwise
896 * leave the filesystem running. We know any error is
897 * local to this block.
899 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
901 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
902 (unsigned long long)bh->b_blocknr);
907 * Errors after here are fatal.
910 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
912 "Extent block #%llu has bad signature %.*s\n",
913 (unsigned long long)bh->b_blocknr, 7,
918 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
920 "Extent block #%llu has an invalid h_blkno of %llu\n",
921 (unsigned long long)bh->b_blocknr,
922 (unsigned long long)le64_to_cpu(eb->h_blkno));
926 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
928 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
929 (unsigned long long)bh->b_blocknr,
930 le32_to_cpu(eb->h_fs_generation));
937 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
938 struct buffer_head **bh)
941 struct buffer_head *tmp = *bh;
943 rc = ocfs2_read_block(ci, eb_blkno, &tmp,
944 ocfs2_validate_extent_block);
946 /* If ocfs2_read_block() got us a new bh, pass it up. */
955 * How many free extents have we got before we need more meta data?
957 int ocfs2_num_free_extents(struct ocfs2_super *osb,
958 struct ocfs2_extent_tree *et)
961 struct ocfs2_extent_list *el = NULL;
962 struct ocfs2_extent_block *eb;
963 struct buffer_head *eb_bh = NULL;
967 last_eb_blk = ocfs2_et_get_last_eb_blk(et);
970 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
976 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
980 BUG_ON(el->l_tree_depth != 0);
982 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
986 trace_ocfs2_num_free_extents(retval);
990 /* expects array to already be allocated
992 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
995 static int ocfs2_create_new_meta_bhs(handle_t *handle,
996 struct ocfs2_extent_tree *et,
998 struct ocfs2_alloc_context *meta_ac,
999 struct buffer_head *bhs[])
1001 int count, status, i;
1002 u16 suballoc_bit_start;
1004 u64 suballoc_loc, first_blkno;
1005 struct ocfs2_super *osb =
1006 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1007 struct ocfs2_extent_block *eb;
1010 while (count < wanted) {
1011 status = ocfs2_claim_metadata(handle,
1015 &suballoc_bit_start,
1023 for(i = count; i < (num_got + count); i++) {
1024 bhs[i] = sb_getblk(osb->sb, first_blkno);
1025 if (bhs[i] == NULL) {
1030 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1032 status = ocfs2_journal_access_eb(handle, et->et_ci,
1034 OCFS2_JOURNAL_ACCESS_CREATE);
1040 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1041 eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1042 /* Ok, setup the minimal stuff here. */
1043 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1044 eb->h_blkno = cpu_to_le64(first_blkno);
1045 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1046 eb->h_suballoc_slot =
1047 cpu_to_le16(meta_ac->ac_alloc_slot);
1048 eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1049 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1050 eb->h_list.l_count =
1051 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1053 suballoc_bit_start++;
1056 /* We'll also be dirtied by the caller, so
1057 * this isn't absolutely necessary. */
1058 ocfs2_journal_dirty(handle, bhs[i]);
1067 for(i = 0; i < wanted; i++) {
1077 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1079 * Returns the sum of the rightmost extent rec logical offset and
1082 * ocfs2_add_branch() uses this to determine what logical cluster
1083 * value should be populated into the leftmost new branch records.
1085 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1086 * value for the new topmost tree record.
1088 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el)
1092 i = le16_to_cpu(el->l_next_free_rec) - 1;
1094 return le32_to_cpu(el->l_recs[i].e_cpos) +
1095 ocfs2_rec_clusters(el, &el->l_recs[i]);
1099 * Change range of the branches in the right most path according to the leaf
1100 * extent block's rightmost record.
1102 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1103 struct ocfs2_extent_tree *et)
1106 struct ocfs2_path *path = NULL;
1107 struct ocfs2_extent_list *el;
1108 struct ocfs2_extent_rec *rec;
1110 path = ocfs2_new_path_from_et(et);
1116 status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1122 status = ocfs2_extend_trans(handle, path_num_items(path));
1128 status = ocfs2_journal_access_path(et->et_ci, handle, path);
1134 el = path_leaf_el(path);
1135 rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1137 ocfs2_adjust_rightmost_records(handle, et, path, rec);
1140 ocfs2_free_path(path);
1145 * Add an entire tree branch to our inode. eb_bh is the extent block
1146 * to start at, if we don't want to start the branch at the root
1149 * last_eb_bh is required as we have to update it's next_leaf pointer
1150 * for the new last extent block.
1152 * the new branch will be 'empty' in the sense that every block will
1153 * contain a single record with cluster count == 0.
1155 static int ocfs2_add_branch(handle_t *handle,
1156 struct ocfs2_extent_tree *et,
1157 struct buffer_head *eb_bh,
1158 struct buffer_head **last_eb_bh,
1159 struct ocfs2_alloc_context *meta_ac)
1161 int status, new_blocks, i;
1162 u64 next_blkno, new_last_eb_blk;
1163 struct buffer_head *bh;
1164 struct buffer_head **new_eb_bhs = NULL;
1165 struct ocfs2_extent_block *eb;
1166 struct ocfs2_extent_list *eb_el;
1167 struct ocfs2_extent_list *el;
1168 u32 new_cpos, root_end;
1170 BUG_ON(!last_eb_bh || !*last_eb_bh);
1173 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1176 el = et->et_root_el;
1178 /* we never add a branch to a leaf. */
1179 BUG_ON(!el->l_tree_depth);
1181 new_blocks = le16_to_cpu(el->l_tree_depth);
1183 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1184 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1185 root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1188 * If there is a gap before the root end and the real end
1189 * of the righmost leaf block, we need to remove the gap
1190 * between new_cpos and root_end first so that the tree
1191 * is consistent after we add a new branch(it will start
1194 if (root_end > new_cpos) {
1195 trace_ocfs2_adjust_rightmost_branch(
1196 (unsigned long long)
1197 ocfs2_metadata_cache_owner(et->et_ci),
1198 root_end, new_cpos);
1200 status = ocfs2_adjust_rightmost_branch(handle, et);
1207 /* allocate the number of new eb blocks we need */
1208 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1216 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1217 meta_ac, new_eb_bhs);
1223 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1224 * linked with the rest of the tree.
1225 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1227 * when we leave the loop, new_last_eb_blk will point to the
1228 * newest leaf, and next_blkno will point to the topmost extent
1230 next_blkno = new_last_eb_blk = 0;
1231 for(i = 0; i < new_blocks; i++) {
1233 eb = (struct ocfs2_extent_block *) bh->b_data;
1234 /* ocfs2_create_new_meta_bhs() should create it right! */
1235 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1236 eb_el = &eb->h_list;
1238 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1239 OCFS2_JOURNAL_ACCESS_CREATE);
1245 eb->h_next_leaf_blk = 0;
1246 eb_el->l_tree_depth = cpu_to_le16(i);
1247 eb_el->l_next_free_rec = cpu_to_le16(1);
1249 * This actually counts as an empty extent as
1252 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1253 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1255 * eb_el isn't always an interior node, but even leaf
1256 * nodes want a zero'd flags and reserved field so
1257 * this gets the whole 32 bits regardless of use.
1259 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1260 if (!eb_el->l_tree_depth)
1261 new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1263 ocfs2_journal_dirty(handle, bh);
1264 next_blkno = le64_to_cpu(eb->h_blkno);
1267 /* This is a bit hairy. We want to update up to three blocks
1268 * here without leaving any of them in an inconsistent state
1269 * in case of error. We don't have to worry about
1270 * journal_dirty erroring as it won't unless we've aborted the
1271 * handle (in which case we would never be here) so reserving
1272 * the write with journal_access is all we need to do. */
1273 status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1274 OCFS2_JOURNAL_ACCESS_WRITE);
1279 status = ocfs2_et_root_journal_access(handle, et,
1280 OCFS2_JOURNAL_ACCESS_WRITE);
1286 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1287 OCFS2_JOURNAL_ACCESS_WRITE);
1294 /* Link the new branch into the rest of the tree (el will
1295 * either be on the root_bh, or the extent block passed in. */
1296 i = le16_to_cpu(el->l_next_free_rec);
1297 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1298 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1299 el->l_recs[i].e_int_clusters = 0;
1300 le16_add_cpu(&el->l_next_free_rec, 1);
1302 /* fe needs a new last extent block pointer, as does the
1303 * next_leaf on the previously last-extent-block. */
1304 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1306 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1307 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1309 ocfs2_journal_dirty(handle, *last_eb_bh);
1310 ocfs2_journal_dirty(handle, et->et_root_bh);
1312 ocfs2_journal_dirty(handle, eb_bh);
1315 * Some callers want to track the rightmost leaf so pass it
1318 brelse(*last_eb_bh);
1319 get_bh(new_eb_bhs[0]);
1320 *last_eb_bh = new_eb_bhs[0];
1325 for (i = 0; i < new_blocks; i++)
1326 brelse(new_eb_bhs[i]);
1334 * adds another level to the allocation tree.
1335 * returns back the new extent block so you can add a branch to it
1338 static int ocfs2_shift_tree_depth(handle_t *handle,
1339 struct ocfs2_extent_tree *et,
1340 struct ocfs2_alloc_context *meta_ac,
1341 struct buffer_head **ret_new_eb_bh)
1345 struct buffer_head *new_eb_bh = NULL;
1346 struct ocfs2_extent_block *eb;
1347 struct ocfs2_extent_list *root_el;
1348 struct ocfs2_extent_list *eb_el;
1350 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1357 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1358 /* ocfs2_create_new_meta_bhs() should create it right! */
1359 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1361 eb_el = &eb->h_list;
1362 root_el = et->et_root_el;
1364 status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1365 OCFS2_JOURNAL_ACCESS_CREATE);
1371 /* copy the root extent list data into the new extent block */
1372 eb_el->l_tree_depth = root_el->l_tree_depth;
1373 eb_el->l_next_free_rec = root_el->l_next_free_rec;
1374 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1375 eb_el->l_recs[i] = root_el->l_recs[i];
1377 ocfs2_journal_dirty(handle, new_eb_bh);
1379 status = ocfs2_et_root_journal_access(handle, et,
1380 OCFS2_JOURNAL_ACCESS_WRITE);
1386 new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1388 /* update root_bh now */
1389 le16_add_cpu(&root_el->l_tree_depth, 1);
1390 root_el->l_recs[0].e_cpos = 0;
1391 root_el->l_recs[0].e_blkno = eb->h_blkno;
1392 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1393 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1394 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1395 root_el->l_next_free_rec = cpu_to_le16(1);
1397 /* If this is our 1st tree depth shift, then last_eb_blk
1398 * becomes the allocated extent block */
1399 if (root_el->l_tree_depth == cpu_to_le16(1))
1400 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1402 ocfs2_journal_dirty(handle, et->et_root_bh);
1404 *ret_new_eb_bh = new_eb_bh;
1414 * Should only be called when there is no space left in any of the
1415 * leaf nodes. What we want to do is find the lowest tree depth
1416 * non-leaf extent block with room for new records. There are three
1417 * valid results of this search:
1419 * 1) a lowest extent block is found, then we pass it back in
1420 * *lowest_eb_bh and return '0'
1422 * 2) the search fails to find anything, but the root_el has room. We
1423 * pass NULL back in *lowest_eb_bh, but still return '0'
1425 * 3) the search fails to find anything AND the root_el is full, in
1426 * which case we return > 0
1428 * return status < 0 indicates an error.
1430 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1431 struct buffer_head **target_bh)
1435 struct ocfs2_extent_block *eb;
1436 struct ocfs2_extent_list *el;
1437 struct buffer_head *bh = NULL;
1438 struct buffer_head *lowest_bh = NULL;
1442 el = et->et_root_el;
1444 while(le16_to_cpu(el->l_tree_depth) > 1) {
1445 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1446 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1447 "Owner %llu has empty extent list (next_free_rec == 0)\n",
1448 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1452 i = le16_to_cpu(el->l_next_free_rec) - 1;
1453 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1455 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1456 "Owner %llu has extent list where extent # %d has no physical block start\n",
1457 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1465 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1471 eb = (struct ocfs2_extent_block *) bh->b_data;
1474 if (le16_to_cpu(el->l_next_free_rec) <
1475 le16_to_cpu(el->l_count)) {
1482 /* If we didn't find one and the fe doesn't have any room,
1483 * then return '1' */
1484 el = et->et_root_el;
1485 if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1488 *target_bh = lowest_bh;
1496 * Grow a b-tree so that it has more records.
1498 * We might shift the tree depth in which case existing paths should
1499 * be considered invalid.
1501 * Tree depth after the grow is returned via *final_depth.
1503 * *last_eb_bh will be updated by ocfs2_add_branch().
1505 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1506 int *final_depth, struct buffer_head **last_eb_bh,
1507 struct ocfs2_alloc_context *meta_ac)
1510 struct ocfs2_extent_list *el = et->et_root_el;
1511 int depth = le16_to_cpu(el->l_tree_depth);
1512 struct buffer_head *bh = NULL;
1514 BUG_ON(meta_ac == NULL);
1516 shift = ocfs2_find_branch_target(et, &bh);
1523 /* We traveled all the way to the bottom of the allocation tree
1524 * and didn't find room for any more extents - we need to add
1525 * another tree level */
1528 trace_ocfs2_grow_tree(
1529 (unsigned long long)
1530 ocfs2_metadata_cache_owner(et->et_ci),
1533 /* ocfs2_shift_tree_depth will return us a buffer with
1534 * the new extent block (so we can pass that to
1535 * ocfs2_add_branch). */
1536 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1544 * Special case: we have room now if we shifted from
1545 * tree_depth 0, so no more work needs to be done.
1547 * We won't be calling add_branch, so pass
1548 * back *last_eb_bh as the new leaf. At depth
1549 * zero, it should always be null so there's
1550 * no reason to brelse.
1552 BUG_ON(*last_eb_bh);
1559 /* call ocfs2_add_branch to add the final part of the tree with
1561 ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1570 *final_depth = depth;
1576 * This function will discard the rightmost extent record.
1578 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1580 int next_free = le16_to_cpu(el->l_next_free_rec);
1581 int count = le16_to_cpu(el->l_count);
1582 unsigned int num_bytes;
1585 /* This will cause us to go off the end of our extent list. */
1586 BUG_ON(next_free >= count);
1588 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1590 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1593 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1594 struct ocfs2_extent_rec *insert_rec)
1596 int i, insert_index, next_free, has_empty, num_bytes;
1597 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1598 struct ocfs2_extent_rec *rec;
1600 next_free = le16_to_cpu(el->l_next_free_rec);
1601 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1605 /* The tree code before us didn't allow enough room in the leaf. */
1606 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1609 * The easiest way to approach this is to just remove the
1610 * empty extent and temporarily decrement next_free.
1614 * If next_free was 1 (only an empty extent), this
1615 * loop won't execute, which is fine. We still want
1616 * the decrement above to happen.
1618 for(i = 0; i < (next_free - 1); i++)
1619 el->l_recs[i] = el->l_recs[i+1];
1625 * Figure out what the new record index should be.
1627 for(i = 0; i < next_free; i++) {
1628 rec = &el->l_recs[i];
1630 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1635 trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1636 has_empty, next_free,
1637 le16_to_cpu(el->l_count));
1639 BUG_ON(insert_index < 0);
1640 BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1641 BUG_ON(insert_index > next_free);
1644 * No need to memmove if we're just adding to the tail.
1646 if (insert_index != next_free) {
1647 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1649 num_bytes = next_free - insert_index;
1650 num_bytes *= sizeof(struct ocfs2_extent_rec);
1651 memmove(&el->l_recs[insert_index + 1],
1652 &el->l_recs[insert_index],
1657 * Either we had an empty extent, and need to re-increment or
1658 * there was no empty extent on a non full rightmost leaf node,
1659 * in which case we still need to increment.
1662 el->l_next_free_rec = cpu_to_le16(next_free);
1664 * Make sure none of the math above just messed up our tree.
1666 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1668 el->l_recs[insert_index] = *insert_rec;
1672 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1674 int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1676 BUG_ON(num_recs == 0);
1678 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1680 size = num_recs * sizeof(struct ocfs2_extent_rec);
1681 memmove(&el->l_recs[0], &el->l_recs[1], size);
1682 memset(&el->l_recs[num_recs], 0,
1683 sizeof(struct ocfs2_extent_rec));
1684 el->l_next_free_rec = cpu_to_le16(num_recs);
1689 * Create an empty extent record .
1691 * l_next_free_rec may be updated.
1693 * If an empty extent already exists do nothing.
1695 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1697 int next_free = le16_to_cpu(el->l_next_free_rec);
1699 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1704 if (ocfs2_is_empty_extent(&el->l_recs[0]))
1707 mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1708 "Asked to create an empty extent in a full list:\n"
1709 "count = %u, tree depth = %u",
1710 le16_to_cpu(el->l_count),
1711 le16_to_cpu(el->l_tree_depth));
1713 ocfs2_shift_records_right(el);
1716 le16_add_cpu(&el->l_next_free_rec, 1);
1717 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1721 * For a rotation which involves two leaf nodes, the "root node" is
1722 * the lowest level tree node which contains a path to both leafs. This
1723 * resulting set of information can be used to form a complete "subtree"
1725 * This function is passed two full paths from the dinode down to a
1726 * pair of adjacent leaves. It's task is to figure out which path
1727 * index contains the subtree root - this can be the root index itself
1728 * in a worst-case rotation.
1730 * The array index of the subtree root is passed back.
1732 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1733 struct ocfs2_path *left,
1734 struct ocfs2_path *right)
1739 * Check that the caller passed in two paths from the same tree.
1741 BUG_ON(path_root_bh(left) != path_root_bh(right));
1747 * The caller didn't pass two adjacent paths.
1749 mlog_bug_on_msg(i > left->p_tree_depth,
1750 "Owner %llu, left depth %u, right depth %u\n"
1751 "left leaf blk %llu, right leaf blk %llu\n",
1752 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1753 left->p_tree_depth, right->p_tree_depth,
1754 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1755 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1756 } while (left->p_node[i].bh->b_blocknr ==
1757 right->p_node[i].bh->b_blocknr);
1762 typedef void (path_insert_t)(void *, struct buffer_head *);
1765 * Traverse a btree path in search of cpos, starting at root_el.
1767 * This code can be called with a cpos larger than the tree, in which
1768 * case it will return the rightmost path.
1770 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1771 struct ocfs2_extent_list *root_el, u32 cpos,
1772 path_insert_t *func, void *data)
1777 struct buffer_head *bh = NULL;
1778 struct ocfs2_extent_block *eb;
1779 struct ocfs2_extent_list *el;
1780 struct ocfs2_extent_rec *rec;
1783 while (el->l_tree_depth) {
1784 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1785 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1786 "Owner %llu has empty extent list at depth %u\n",
1787 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1788 le16_to_cpu(el->l_tree_depth));
1794 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1795 rec = &el->l_recs[i];
1798 * In the case that cpos is off the allocation
1799 * tree, this should just wind up returning the
1802 range = le32_to_cpu(rec->e_cpos) +
1803 ocfs2_rec_clusters(el, rec);
1804 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1808 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1810 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1811 "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1812 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1813 le16_to_cpu(el->l_tree_depth), i);
1820 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1826 eb = (struct ocfs2_extent_block *) bh->b_data;
1829 if (le16_to_cpu(el->l_next_free_rec) >
1830 le16_to_cpu(el->l_count)) {
1831 ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1832 "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1833 (unsigned long long)ocfs2_metadata_cache_owner(ci),
1834 (unsigned long long)bh->b_blocknr,
1835 le16_to_cpu(el->l_next_free_rec),
1836 le16_to_cpu(el->l_count));
1847 * Catch any trailing bh that the loop didn't handle.
1855 * Given an initialized path (that is, it has a valid root extent
1856 * list), this function will traverse the btree in search of the path
1857 * which would contain cpos.
1859 * The path traveled is recorded in the path structure.
1861 * Note that this will not do any comparisons on leaf node extent
1862 * records, so it will work fine in the case that we just added a tree
1865 struct find_path_data {
1867 struct ocfs2_path *path;
1869 static void find_path_ins(void *data, struct buffer_head *bh)
1871 struct find_path_data *fp = data;
1874 ocfs2_path_insert_eb(fp->path, fp->index, bh);
1877 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1878 struct ocfs2_path *path, u32 cpos)
1880 struct find_path_data data;
1884 return __ocfs2_find_path(ci, path_root_el(path), cpos,
1885 find_path_ins, &data);
1888 static void find_leaf_ins(void *data, struct buffer_head *bh)
1890 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1891 struct ocfs2_extent_list *el = &eb->h_list;
1892 struct buffer_head **ret = data;
1894 /* We want to retain only the leaf block. */
1895 if (le16_to_cpu(el->l_tree_depth) == 0) {
1901 * Find the leaf block in the tree which would contain cpos. No
1902 * checking of the actual leaf is done.
1904 * Some paths want to call this instead of allocating a path structure
1905 * and calling ocfs2_find_path().
1907 * This function doesn't handle non btree extent lists.
1909 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1910 struct ocfs2_extent_list *root_el, u32 cpos,
1911 struct buffer_head **leaf_bh)
1914 struct buffer_head *bh = NULL;
1916 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1928 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1930 * Basically, we've moved stuff around at the bottom of the tree and
1931 * we need to fix up the extent records above the changes to reflect
1934 * left_rec: the record on the left.
1935 * left_child_el: is the child list pointed to by left_rec
1936 * right_rec: the record to the right of left_rec
1937 * right_child_el: is the child list pointed to by right_rec
1939 * By definition, this only works on interior nodes.
1941 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1942 struct ocfs2_extent_list *left_child_el,
1943 struct ocfs2_extent_rec *right_rec,
1944 struct ocfs2_extent_list *right_child_el)
1946 u32 left_clusters, right_end;
1949 * Interior nodes never have holes. Their cpos is the cpos of
1950 * the leftmost record in their child list. Their cluster
1951 * count covers the full theoretical range of their child list
1952 * - the range between their cpos and the cpos of the record
1953 * immediately to their right.
1955 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1956 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1957 BUG_ON(right_child_el->l_tree_depth);
1958 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1959 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1961 left_clusters -= le32_to_cpu(left_rec->e_cpos);
1962 left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1965 * Calculate the rightmost cluster count boundary before
1966 * moving cpos - we will need to adjust clusters after
1967 * updating e_cpos to keep the same highest cluster count.
1969 right_end = le32_to_cpu(right_rec->e_cpos);
1970 right_end += le32_to_cpu(right_rec->e_int_clusters);
1972 right_rec->e_cpos = left_rec->e_cpos;
1973 le32_add_cpu(&right_rec->e_cpos, left_clusters);
1975 right_end -= le32_to_cpu(right_rec->e_cpos);
1976 right_rec->e_int_clusters = cpu_to_le32(right_end);
1980 * Adjust the adjacent root node records involved in a
1981 * rotation. left_el_blkno is passed in as a key so that we can easily
1982 * find it's index in the root list.
1984 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1985 struct ocfs2_extent_list *left_el,
1986 struct ocfs2_extent_list *right_el,
1991 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
1992 le16_to_cpu(left_el->l_tree_depth));
1994 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
1995 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2000 * The path walking code should have never returned a root and
2001 * two paths which are not adjacent.
2003 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2005 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2006 &root_el->l_recs[i + 1], right_el);
2010 * We've changed a leaf block (in right_path) and need to reflect that
2011 * change back up the subtree.
2013 * This happens in multiple places:
2014 * - When we've moved an extent record from the left path leaf to the right
2015 * path leaf to make room for an empty extent in the left path leaf.
2016 * - When our insert into the right path leaf is at the leftmost edge
2017 * and requires an update of the path immediately to it's left. This
2018 * can occur at the end of some types of rotation and appending inserts.
2019 * - When we've adjusted the last extent record in the left path leaf and the
2020 * 1st extent record in the right path leaf during cross extent block merge.
2022 static void ocfs2_complete_edge_insert(handle_t *handle,
2023 struct ocfs2_path *left_path,
2024 struct ocfs2_path *right_path,
2028 struct ocfs2_extent_list *el, *left_el, *right_el;
2029 struct ocfs2_extent_rec *left_rec, *right_rec;
2030 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2033 * Update the counts and position values within all the
2034 * interior nodes to reflect the leaf rotation we just did.
2036 * The root node is handled below the loop.
2038 * We begin the loop with right_el and left_el pointing to the
2039 * leaf lists and work our way up.
2041 * NOTE: within this loop, left_el and right_el always refer
2042 * to the *child* lists.
2044 left_el = path_leaf_el(left_path);
2045 right_el = path_leaf_el(right_path);
2046 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2047 trace_ocfs2_complete_edge_insert(i);
2050 * One nice property of knowing that all of these
2051 * nodes are below the root is that we only deal with
2052 * the leftmost right node record and the rightmost
2055 el = left_path->p_node[i].el;
2056 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2057 left_rec = &el->l_recs[idx];
2059 el = right_path->p_node[i].el;
2060 right_rec = &el->l_recs[0];
2062 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2065 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2066 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2069 * Setup our list pointers now so that the current
2070 * parents become children in the next iteration.
2072 left_el = left_path->p_node[i].el;
2073 right_el = right_path->p_node[i].el;
2077 * At the root node, adjust the two adjacent records which
2078 * begin our path to the leaves.
2081 el = left_path->p_node[subtree_index].el;
2082 left_el = left_path->p_node[subtree_index + 1].el;
2083 right_el = right_path->p_node[subtree_index + 1].el;
2085 ocfs2_adjust_root_records(el, left_el, right_el,
2086 left_path->p_node[subtree_index + 1].bh->b_blocknr);
2088 root_bh = left_path->p_node[subtree_index].bh;
2090 ocfs2_journal_dirty(handle, root_bh);
2093 static int ocfs2_rotate_subtree_right(handle_t *handle,
2094 struct ocfs2_extent_tree *et,
2095 struct ocfs2_path *left_path,
2096 struct ocfs2_path *right_path,
2100 struct buffer_head *right_leaf_bh;
2101 struct buffer_head *left_leaf_bh = NULL;
2102 struct buffer_head *root_bh;
2103 struct ocfs2_extent_list *right_el, *left_el;
2104 struct ocfs2_extent_rec move_rec;
2106 left_leaf_bh = path_leaf_bh(left_path);
2107 left_el = path_leaf_el(left_path);
2109 if (left_el->l_next_free_rec != left_el->l_count) {
2110 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2111 "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2112 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2113 (unsigned long long)left_leaf_bh->b_blocknr,
2114 le16_to_cpu(left_el->l_next_free_rec));
2119 * This extent block may already have an empty record, so we
2120 * return early if so.
2122 if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2125 root_bh = left_path->p_node[subtree_index].bh;
2126 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2128 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2135 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2136 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2143 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2151 right_leaf_bh = path_leaf_bh(right_path);
2152 right_el = path_leaf_el(right_path);
2154 /* This is a code error, not a disk corruption. */
2155 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2156 "because rightmost leaf block %llu is empty\n",
2157 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2158 (unsigned long long)right_leaf_bh->b_blocknr);
2160 ocfs2_create_empty_extent(right_el);
2162 ocfs2_journal_dirty(handle, right_leaf_bh);
2164 /* Do the copy now. */
2165 i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2166 move_rec = left_el->l_recs[i];
2167 right_el->l_recs[0] = move_rec;
2170 * Clear out the record we just copied and shift everything
2171 * over, leaving an empty extent in the left leaf.
2173 * We temporarily subtract from next_free_rec so that the
2174 * shift will lose the tail record (which is now defunct).
2176 le16_add_cpu(&left_el->l_next_free_rec, -1);
2177 ocfs2_shift_records_right(left_el);
2178 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2179 le16_add_cpu(&left_el->l_next_free_rec, 1);
2181 ocfs2_journal_dirty(handle, left_leaf_bh);
2183 ocfs2_complete_edge_insert(handle, left_path, right_path,
2191 * Given a full path, determine what cpos value would return us a path
2192 * containing the leaf immediately to the left of the current one.
2194 * Will return zero if the path passed in is already the leftmost path.
2196 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2197 struct ocfs2_path *path, u32 *cpos)
2201 struct ocfs2_extent_list *el;
2203 BUG_ON(path->p_tree_depth == 0);
2207 blkno = path_leaf_bh(path)->b_blocknr;
2209 /* Start at the tree node just above the leaf and work our way up. */
2210 i = path->p_tree_depth - 1;
2212 el = path->p_node[i].el;
2215 * Find the extent record just before the one in our
2218 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2219 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2223 * We've determined that the
2224 * path specified is already
2225 * the leftmost one - return a
2231 * The leftmost record points to our
2232 * leaf - we need to travel up the
2238 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2239 *cpos = *cpos + ocfs2_rec_clusters(el,
2240 &el->l_recs[j - 1]);
2247 * If we got here, we never found a valid node where
2248 * the tree indicated one should be.
2250 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2251 (unsigned long long)blkno);
2256 blkno = path->p_node[i].bh->b_blocknr;
2265 * Extend the transaction by enough credits to complete the rotation,
2266 * and still leave at least the original number of credits allocated
2267 * to this transaction.
2269 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2271 struct ocfs2_path *path)
2274 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2276 if (handle->h_buffer_credits < credits)
2277 ret = ocfs2_extend_trans(handle,
2278 credits - handle->h_buffer_credits);
2284 * Trap the case where we're inserting into the theoretical range past
2285 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2286 * whose cpos is less than ours into the right leaf.
2288 * It's only necessary to look at the rightmost record of the left
2289 * leaf because the logic that calls us should ensure that the
2290 * theoretical ranges in the path components above the leaves are
2293 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2296 struct ocfs2_extent_list *left_el;
2297 struct ocfs2_extent_rec *rec;
2300 left_el = path_leaf_el(left_path);
2301 next_free = le16_to_cpu(left_el->l_next_free_rec);
2302 rec = &left_el->l_recs[next_free - 1];
2304 if (insert_cpos > le32_to_cpu(rec->e_cpos))
2309 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2311 int next_free = le16_to_cpu(el->l_next_free_rec);
2313 struct ocfs2_extent_rec *rec;
2318 rec = &el->l_recs[0];
2319 if (ocfs2_is_empty_extent(rec)) {
2323 rec = &el->l_recs[1];
2326 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2327 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2333 * Rotate all the records in a btree right one record, starting at insert_cpos.
2335 * The path to the rightmost leaf should be passed in.
2337 * The array is assumed to be large enough to hold an entire path (tree depth).
2339 * Upon successful return from this function:
2341 * - The 'right_path' array will contain a path to the leaf block
2342 * whose range contains e_cpos.
2343 * - That leaf block will have a single empty extent in list index 0.
2344 * - In the case that the rotation requires a post-insert update,
2345 * *ret_left_path will contain a valid path which can be passed to
2346 * ocfs2_insert_path().
2348 static int ocfs2_rotate_tree_right(handle_t *handle,
2349 struct ocfs2_extent_tree *et,
2350 enum ocfs2_split_type split,
2352 struct ocfs2_path *right_path,
2353 struct ocfs2_path **ret_left_path)
2355 int ret, start, orig_credits = handle->h_buffer_credits;
2357 struct ocfs2_path *left_path = NULL;
2358 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2360 *ret_left_path = NULL;
2362 left_path = ocfs2_new_path_from_path(right_path);
2369 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2375 trace_ocfs2_rotate_tree_right(
2376 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2380 * What we want to do here is:
2382 * 1) Start with the rightmost path.
2384 * 2) Determine a path to the leaf block directly to the left
2387 * 3) Determine the 'subtree root' - the lowest level tree node
2388 * which contains a path to both leaves.
2390 * 4) Rotate the subtree.
2392 * 5) Find the next subtree by considering the left path to be
2393 * the new right path.
2395 * The check at the top of this while loop also accepts
2396 * insert_cpos == cpos because cpos is only a _theoretical_
2397 * value to get us the left path - insert_cpos might very well
2398 * be filling that hole.
2400 * Stop at a cpos of '0' because we either started at the
2401 * leftmost branch (i.e., a tree with one branch and a
2402 * rotation inside of it), or we've gone as far as we can in
2403 * rotating subtrees.
2405 while (cpos && insert_cpos <= cpos) {
2406 trace_ocfs2_rotate_tree_right(
2407 (unsigned long long)
2408 ocfs2_metadata_cache_owner(et->et_ci),
2411 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2417 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2418 path_leaf_bh(right_path),
2419 "Owner %llu: error during insert of %u "
2420 "(left path cpos %u) results in two identical "
2421 "paths ending at %llu\n",
2422 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2424 (unsigned long long)
2425 path_leaf_bh(left_path)->b_blocknr);
2427 if (split == SPLIT_NONE &&
2428 ocfs2_rotate_requires_path_adjustment(left_path,
2432 * We've rotated the tree as much as we
2433 * should. The rest is up to
2434 * ocfs2_insert_path() to complete, after the
2435 * record insertion. We indicate this
2436 * situation by returning the left path.
2438 * The reason we don't adjust the records here
2439 * before the record insert is that an error
2440 * later might break the rule where a parent
2441 * record e_cpos will reflect the actual
2442 * e_cpos of the 1st nonempty record of the
2445 *ret_left_path = left_path;
2449 start = ocfs2_find_subtree_root(et, left_path, right_path);
2451 trace_ocfs2_rotate_subtree(start,
2452 (unsigned long long)
2453 right_path->p_node[start].bh->b_blocknr,
2454 right_path->p_tree_depth);
2456 ret = ocfs2_extend_rotate_transaction(handle, start,
2457 orig_credits, right_path);
2463 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2470 if (split != SPLIT_NONE &&
2471 ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2474 * A rotate moves the rightmost left leaf
2475 * record over to the leftmost right leaf
2476 * slot. If we're doing an extent split
2477 * instead of a real insert, then we have to
2478 * check that the extent to be split wasn't
2479 * just moved over. If it was, then we can
2480 * exit here, passing left_path back -
2481 * ocfs2_split_extent() is smart enough to
2482 * search both leaves.
2484 *ret_left_path = left_path;
2489 * There is no need to re-read the next right path
2490 * as we know that it'll be our current left
2491 * path. Optimize by copying values instead.
2493 ocfs2_mv_path(right_path, left_path);
2495 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2503 ocfs2_free_path(left_path);
2509 static int ocfs2_update_edge_lengths(handle_t *handle,
2510 struct ocfs2_extent_tree *et,
2511 int subtree_index, struct ocfs2_path *path)
2514 struct ocfs2_extent_rec *rec;
2515 struct ocfs2_extent_list *el;
2516 struct ocfs2_extent_block *eb;
2520 * In normal tree rotation process, we will never touch the
2521 * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2522 * doesn't reserve the credits for them either.
2524 * But we do have a special case here which will update the rightmost
2525 * records for all the bh in the path.
2526 * So we have to allocate extra credits and access them.
2528 ret = ocfs2_extend_trans(handle, subtree_index);
2534 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2540 /* Path should always be rightmost. */
2541 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2542 BUG_ON(eb->h_next_leaf_blk != 0ULL);
2545 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2546 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2547 rec = &el->l_recs[idx];
2548 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2550 for (i = 0; i < path->p_tree_depth; i++) {
2551 el = path->p_node[i].el;
2552 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2553 rec = &el->l_recs[idx];
2555 rec->e_int_clusters = cpu_to_le32(range);
2556 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2558 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2564 static void ocfs2_unlink_path(handle_t *handle,
2565 struct ocfs2_extent_tree *et,
2566 struct ocfs2_cached_dealloc_ctxt *dealloc,
2567 struct ocfs2_path *path, int unlink_start)
2570 struct ocfs2_extent_block *eb;
2571 struct ocfs2_extent_list *el;
2572 struct buffer_head *bh;
2574 for(i = unlink_start; i < path_num_items(path); i++) {
2575 bh = path->p_node[i].bh;
2577 eb = (struct ocfs2_extent_block *)bh->b_data;
2579 * Not all nodes might have had their final count
2580 * decremented by the caller - handle this here.
2583 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2585 "Inode %llu, attempted to remove extent block "
2586 "%llu with %u records\n",
2587 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2588 (unsigned long long)le64_to_cpu(eb->h_blkno),
2589 le16_to_cpu(el->l_next_free_rec));
2591 ocfs2_journal_dirty(handle, bh);
2592 ocfs2_remove_from_cache(et->et_ci, bh);
2596 el->l_next_free_rec = 0;
2597 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2599 ocfs2_journal_dirty(handle, bh);
2601 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2605 ocfs2_remove_from_cache(et->et_ci, bh);
2609 static void ocfs2_unlink_subtree(handle_t *handle,
2610 struct ocfs2_extent_tree *et,
2611 struct ocfs2_path *left_path,
2612 struct ocfs2_path *right_path,
2614 struct ocfs2_cached_dealloc_ctxt *dealloc)
2617 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2618 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2619 struct ocfs2_extent_list *el;
2620 struct ocfs2_extent_block *eb;
2622 el = path_leaf_el(left_path);
2624 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2626 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2627 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2630 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2632 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2633 le16_add_cpu(&root_el->l_next_free_rec, -1);
2635 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2636 eb->h_next_leaf_blk = 0;
2638 ocfs2_journal_dirty(handle, root_bh);
2639 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2641 ocfs2_unlink_path(handle, et, dealloc, right_path,
2645 static int ocfs2_rotate_subtree_left(handle_t *handle,
2646 struct ocfs2_extent_tree *et,
2647 struct ocfs2_path *left_path,
2648 struct ocfs2_path *right_path,
2650 struct ocfs2_cached_dealloc_ctxt *dealloc,
2653 int ret, i, del_right_subtree = 0, right_has_empty = 0;
2654 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2655 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2656 struct ocfs2_extent_block *eb;
2660 right_leaf_el = path_leaf_el(right_path);
2661 left_leaf_el = path_leaf_el(left_path);
2662 root_bh = left_path->p_node[subtree_index].bh;
2663 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2665 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2668 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2669 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2671 * It's legal for us to proceed if the right leaf is
2672 * the rightmost one and it has an empty extent. There
2673 * are two cases to handle - whether the leaf will be
2674 * empty after removal or not. If the leaf isn't empty
2675 * then just remove the empty extent up front. The
2676 * next block will handle empty leaves by flagging
2679 * Non rightmost leaves will throw -EAGAIN and the
2680 * caller can manually move the subtree and retry.
2683 if (eb->h_next_leaf_blk != 0ULL)
2686 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2687 ret = ocfs2_journal_access_eb(handle, et->et_ci,
2688 path_leaf_bh(right_path),
2689 OCFS2_JOURNAL_ACCESS_WRITE);
2695 ocfs2_remove_empty_extent(right_leaf_el);
2697 right_has_empty = 1;
2700 if (eb->h_next_leaf_blk == 0ULL &&
2701 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2703 * We have to update i_last_eb_blk during the meta
2706 ret = ocfs2_et_root_journal_access(handle, et,
2707 OCFS2_JOURNAL_ACCESS_WRITE);
2713 del_right_subtree = 1;
2717 * Getting here with an empty extent in the right path implies
2718 * that it's the rightmost path and will be deleted.
2720 BUG_ON(right_has_empty && !del_right_subtree);
2722 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2729 for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2730 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2737 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2745 if (!right_has_empty) {
2747 * Only do this if we're moving a real
2748 * record. Otherwise, the action is delayed until
2749 * after removal of the right path in which case we
2750 * can do a simple shift to remove the empty extent.
2752 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2753 memset(&right_leaf_el->l_recs[0], 0,
2754 sizeof(struct ocfs2_extent_rec));
2756 if (eb->h_next_leaf_blk == 0ULL) {
2758 * Move recs over to get rid of empty extent, decrease
2759 * next_free. This is allowed to remove the last
2760 * extent in our leaf (setting l_next_free_rec to
2761 * zero) - the delete code below won't care.
2763 ocfs2_remove_empty_extent(right_leaf_el);
2766 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2767 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2769 if (del_right_subtree) {
2770 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2771 subtree_index, dealloc);
2772 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2779 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2780 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2783 * Removal of the extent in the left leaf was skipped
2784 * above so we could delete the right path
2787 if (right_has_empty)
2788 ocfs2_remove_empty_extent(left_leaf_el);
2790 ocfs2_journal_dirty(handle, et_root_bh);
2794 ocfs2_complete_edge_insert(handle, left_path, right_path,
2802 * Given a full path, determine what cpos value would return us a path
2803 * containing the leaf immediately to the right of the current one.
2805 * Will return zero if the path passed in is already the rightmost path.
2807 * This looks similar, but is subtly different to
2808 * ocfs2_find_cpos_for_left_leaf().
2810 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2811 struct ocfs2_path *path, u32 *cpos)
2815 struct ocfs2_extent_list *el;
2819 if (path->p_tree_depth == 0)
2822 blkno = path_leaf_bh(path)->b_blocknr;
2824 /* Start at the tree node just above the leaf and work our way up. */
2825 i = path->p_tree_depth - 1;
2829 el = path->p_node[i].el;
2832 * Find the extent record just after the one in our
2835 next_free = le16_to_cpu(el->l_next_free_rec);
2836 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2837 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2838 if (j == (next_free - 1)) {
2841 * We've determined that the
2842 * path specified is already
2843 * the rightmost one - return a
2849 * The rightmost record points to our
2850 * leaf - we need to travel up the
2856 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2862 * If we got here, we never found a valid node where
2863 * the tree indicated one should be.
2865 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2866 (unsigned long long)blkno);
2871 blkno = path->p_node[i].bh->b_blocknr;
2879 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2880 struct ocfs2_extent_tree *et,
2881 struct ocfs2_path *path)
2884 struct buffer_head *bh = path_leaf_bh(path);
2885 struct ocfs2_extent_list *el = path_leaf_el(path);
2887 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2890 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2891 path_num_items(path) - 1);
2897 ocfs2_remove_empty_extent(el);
2898 ocfs2_journal_dirty(handle, bh);
2904 static int __ocfs2_rotate_tree_left(handle_t *handle,
2905 struct ocfs2_extent_tree *et,
2907 struct ocfs2_path *path,
2908 struct ocfs2_cached_dealloc_ctxt *dealloc,
2909 struct ocfs2_path **empty_extent_path)
2911 int ret, subtree_root, deleted;
2913 struct ocfs2_path *left_path = NULL;
2914 struct ocfs2_path *right_path = NULL;
2915 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2917 if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2920 *empty_extent_path = NULL;
2922 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2928 left_path = ocfs2_new_path_from_path(path);
2935 ocfs2_cp_path(left_path, path);
2937 right_path = ocfs2_new_path_from_path(path);
2944 while (right_cpos) {
2945 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2951 subtree_root = ocfs2_find_subtree_root(et, left_path,
2954 trace_ocfs2_rotate_subtree(subtree_root,
2955 (unsigned long long)
2956 right_path->p_node[subtree_root].bh->b_blocknr,
2957 right_path->p_tree_depth);
2959 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2960 orig_credits, left_path);
2967 * Caller might still want to make changes to the
2968 * tree root, so re-add it to the journal here.
2970 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2977 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2978 right_path, subtree_root,
2980 if (ret == -EAGAIN) {
2982 * The rotation has to temporarily stop due to
2983 * the right subtree having an empty
2984 * extent. Pass it back to the caller for a
2987 *empty_extent_path = right_path;
2997 * The subtree rotate might have removed records on
2998 * the rightmost edge. If so, then rotation is
3004 ocfs2_mv_path(left_path, right_path);
3006 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3015 ocfs2_free_path(right_path);
3016 ocfs2_free_path(left_path);
3021 static int ocfs2_remove_rightmost_path(handle_t *handle,
3022 struct ocfs2_extent_tree *et,
3023 struct ocfs2_path *path,
3024 struct ocfs2_cached_dealloc_ctxt *dealloc)
3026 int ret, subtree_index;
3028 struct ocfs2_path *left_path = NULL;
3029 struct ocfs2_extent_block *eb;
3030 struct ocfs2_extent_list *el;
3033 ret = ocfs2_et_sanity_check(et);
3037 * There's two ways we handle this depending on
3038 * whether path is the only existing one.
3040 ret = ocfs2_extend_rotate_transaction(handle, 0,
3041 handle->h_buffer_credits,
3048 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3054 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3063 * We have a path to the left of this one - it needs
3066 left_path = ocfs2_new_path_from_path(path);
3073 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3079 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3085 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3087 ocfs2_unlink_subtree(handle, et, left_path, path,
3088 subtree_index, dealloc);
3089 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3096 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3097 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3100 * 'path' is also the leftmost path which
3101 * means it must be the only one. This gets
3102 * handled differently because we want to
3103 * revert the root back to having extents
3106 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3108 el = et->et_root_el;
3109 el->l_tree_depth = 0;
3110 el->l_next_free_rec = 0;
3111 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3113 ocfs2_et_set_last_eb_blk(et, 0);
3116 ocfs2_journal_dirty(handle, path_root_bh(path));
3119 ocfs2_free_path(left_path);
3123 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3124 struct ocfs2_extent_tree *et,
3125 struct ocfs2_path *path,
3126 struct ocfs2_cached_dealloc_ctxt *dealloc)
3130 int credits = path->p_tree_depth * 2 + 1;
3132 handle = ocfs2_start_trans(osb, credits);
3133 if (IS_ERR(handle)) {
3134 ret = PTR_ERR(handle);
3139 ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3143 ocfs2_commit_trans(osb, handle);
3148 * Left rotation of btree records.
3150 * In many ways, this is (unsurprisingly) the opposite of right
3151 * rotation. We start at some non-rightmost path containing an empty
3152 * extent in the leaf block. The code works its way to the rightmost
3153 * path by rotating records to the left in every subtree.
3155 * This is used by any code which reduces the number of extent records
3156 * in a leaf. After removal, an empty record should be placed in the
3157 * leftmost list position.
3159 * This won't handle a length update of the rightmost path records if
3160 * the rightmost tree leaf record is removed so the caller is
3161 * responsible for detecting and correcting that.
3163 static int ocfs2_rotate_tree_left(handle_t *handle,
3164 struct ocfs2_extent_tree *et,
3165 struct ocfs2_path *path,
3166 struct ocfs2_cached_dealloc_ctxt *dealloc)
3168 int ret, orig_credits = handle->h_buffer_credits;
3169 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3170 struct ocfs2_extent_block *eb;
3171 struct ocfs2_extent_list *el;
3173 el = path_leaf_el(path);
3174 if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3177 if (path->p_tree_depth == 0) {
3178 rightmost_no_delete:
3180 * Inline extents. This is trivially handled, so do
3183 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3190 * Handle rightmost branch now. There's several cases:
3191 * 1) simple rotation leaving records in there. That's trivial.
3192 * 2) rotation requiring a branch delete - there's no more
3193 * records left. Two cases of this:
3194 * a) There are branches to the left.
3195 * b) This is also the leftmost (the only) branch.
3197 * 1) is handled via ocfs2_rotate_rightmost_leaf_left()
3198 * 2a) we need the left branch so that we can update it with the unlink
3199 * 2b) we need to bring the root back to inline extents.
3202 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3204 if (eb->h_next_leaf_blk == 0) {
3206 * This gets a bit tricky if we're going to delete the
3207 * rightmost path. Get the other cases out of the way
3210 if (le16_to_cpu(el->l_next_free_rec) > 1)
3211 goto rightmost_no_delete;
3213 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3215 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3216 "Owner %llu has empty extent block at %llu\n",
3217 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3218 (unsigned long long)le64_to_cpu(eb->h_blkno));
3223 * XXX: The caller can not trust "path" any more after
3224 * this as it will have been deleted. What do we do?
3226 * In theory the rotate-for-merge code will never get
3227 * here because it'll always ask for a rotate in a
3231 ret = ocfs2_remove_rightmost_path(handle, et, path,
3239 * Now we can loop, remembering the path we get from -EAGAIN
3240 * and restarting from there.
3243 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3244 dealloc, &restart_path);
3245 if (ret && ret != -EAGAIN) {
3250 while (ret == -EAGAIN) {
3251 tmp_path = restart_path;
3252 restart_path = NULL;
3254 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3257 if (ret && ret != -EAGAIN) {
3262 ocfs2_free_path(tmp_path);
3270 ocfs2_free_path(tmp_path);
3271 ocfs2_free_path(restart_path);
3275 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3278 struct ocfs2_extent_rec *rec = &el->l_recs[index];
3281 if (rec->e_leaf_clusters == 0) {
3283 * We consumed all of the merged-from record. An empty
3284 * extent cannot exist anywhere but the 1st array
3285 * position, so move things over if the merged-from
3286 * record doesn't occupy that position.
3288 * This creates a new empty extent so the caller
3289 * should be smart enough to have removed any existing
3293 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3294 size = index * sizeof(struct ocfs2_extent_rec);
3295 memmove(&el->l_recs[1], &el->l_recs[0], size);
3299 * Always memset - the caller doesn't check whether it
3300 * created an empty extent, so there could be junk in
3303 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3307 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3308 struct ocfs2_path *left_path,
3309 struct ocfs2_path **ret_right_path)
3313 struct ocfs2_path *right_path = NULL;
3314 struct ocfs2_extent_list *left_el;
3316 *ret_right_path = NULL;
3318 /* This function shouldn't be called for non-trees. */
3319 BUG_ON(left_path->p_tree_depth == 0);
3321 left_el = path_leaf_el(left_path);
3322 BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3324 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3325 left_path, &right_cpos);
3331 /* This function shouldn't be called for the rightmost leaf. */
3332 BUG_ON(right_cpos == 0);
3334 right_path = ocfs2_new_path_from_path(left_path);
3341 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3347 *ret_right_path = right_path;
3350 ocfs2_free_path(right_path);
3355 * Remove split_rec clusters from the record at index and merge them
3356 * onto the beginning of the record "next" to it.
3357 * For index < l_count - 1, the next means the extent rec at index + 1.
3358 * For index == l_count - 1, the "next" means the 1st extent rec of the
3359 * next extent block.
3361 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3363 struct ocfs2_extent_tree *et,
3364 struct ocfs2_extent_rec *split_rec,
3367 int ret, next_free, i;
3368 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3369 struct ocfs2_extent_rec *left_rec;
3370 struct ocfs2_extent_rec *right_rec;
3371 struct ocfs2_extent_list *right_el;
3372 struct ocfs2_path *right_path = NULL;
3373 int subtree_index = 0;
3374 struct ocfs2_extent_list *el = path_leaf_el(left_path);
3375 struct buffer_head *bh = path_leaf_bh(left_path);
3376 struct buffer_head *root_bh = NULL;
3378 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3379 left_rec = &el->l_recs[index];
3381 if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3382 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3383 /* we meet with a cross extent block merge. */
3384 ret = ocfs2_get_right_path(et, left_path, &right_path);
3390 right_el = path_leaf_el(right_path);
3391 next_free = le16_to_cpu(right_el->l_next_free_rec);
3392 BUG_ON(next_free <= 0);
3393 right_rec = &right_el->l_recs[0];
3394 if (ocfs2_is_empty_extent(right_rec)) {
3395 BUG_ON(next_free <= 1);
3396 right_rec = &right_el->l_recs[1];
3399 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3400 le16_to_cpu(left_rec->e_leaf_clusters) !=
3401 le32_to_cpu(right_rec->e_cpos));
3403 subtree_index = ocfs2_find_subtree_root(et, left_path,
3406 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3407 handle->h_buffer_credits,
3414 root_bh = left_path->p_node[subtree_index].bh;
3415 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3417 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3424 for (i = subtree_index + 1;
3425 i < path_num_items(right_path); i++) {
3426 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3433 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3442 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3443 right_rec = &el->l_recs[index + 1];
3446 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3447 path_num_items(left_path) - 1);
3453 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3455 le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3456 le64_add_cpu(&right_rec->e_blkno,
3457 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3459 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3461 ocfs2_cleanup_merge(el, index);
3463 ocfs2_journal_dirty(handle, bh);
3465 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3466 ocfs2_complete_edge_insert(handle, left_path, right_path,
3470 ocfs2_free_path(right_path);
3474 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3475 struct ocfs2_path *right_path,
3476 struct ocfs2_path **ret_left_path)
3480 struct ocfs2_path *left_path = NULL;
3482 *ret_left_path = NULL;
3484 /* This function shouldn't be called for non-trees. */
3485 BUG_ON(right_path->p_tree_depth == 0);
3487 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3488 right_path, &left_cpos);
3494 /* This function shouldn't be called for the leftmost leaf. */
3495 BUG_ON(left_cpos == 0);
3497 left_path = ocfs2_new_path_from_path(right_path);
3504 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3510 *ret_left_path = left_path;
3513 ocfs2_free_path(left_path);
3518 * Remove split_rec clusters from the record at index and merge them
3519 * onto the tail of the record "before" it.
3520 * For index > 0, the "before" means the extent rec at index - 1.
3522 * For index == 0, the "before" means the last record of the previous
3523 * extent block. And there is also a situation that we may need to
3524 * remove the rightmost leaf extent block in the right_path and change
3525 * the right path to indicate the new rightmost path.
3527 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3529 struct ocfs2_extent_tree *et,
3530 struct ocfs2_extent_rec *split_rec,
3531 struct ocfs2_cached_dealloc_ctxt *dealloc,
3534 int ret, i, subtree_index = 0, has_empty_extent = 0;
3535 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3536 struct ocfs2_extent_rec *left_rec;
3537 struct ocfs2_extent_rec *right_rec;
3538 struct ocfs2_extent_list *el = path_leaf_el(right_path);
3539 struct buffer_head *bh = path_leaf_bh(right_path);
3540 struct buffer_head *root_bh = NULL;
3541 struct ocfs2_path *left_path = NULL;
3542 struct ocfs2_extent_list *left_el;
3546 right_rec = &el->l_recs[index];
3548 /* we meet with a cross extent block merge. */
3549 ret = ocfs2_get_left_path(et, right_path, &left_path);
3555 left_el = path_leaf_el(left_path);
3556 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3557 le16_to_cpu(left_el->l_count));
3559 left_rec = &left_el->l_recs[
3560 le16_to_cpu(left_el->l_next_free_rec) - 1];
3561 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3562 le16_to_cpu(left_rec->e_leaf_clusters) !=
3563 le32_to_cpu(split_rec->e_cpos));
3565 subtree_index = ocfs2_find_subtree_root(et, left_path,
3568 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3569 handle->h_buffer_credits,
3576 root_bh = left_path->p_node[subtree_index].bh;
3577 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3579 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3586 for (i = subtree_index + 1;
3587 i < path_num_items(right_path); i++) {
3588 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3595 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3603 left_rec = &el->l_recs[index - 1];
3604 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3605 has_empty_extent = 1;
3608 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3609 path_num_items(right_path) - 1);
3615 if (has_empty_extent && index == 1) {
3617 * The easy case - we can just plop the record right in.
3619 *left_rec = *split_rec;
3621 has_empty_extent = 0;
3623 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3625 le32_add_cpu(&right_rec->e_cpos, split_clusters);
3626 le64_add_cpu(&right_rec->e_blkno,
3627 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3629 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3631 ocfs2_cleanup_merge(el, index);
3633 ocfs2_journal_dirty(handle, bh);
3635 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3638 * In the situation that the right_rec is empty and the extent
3639 * block is empty also, ocfs2_complete_edge_insert can't handle
3640 * it and we need to delete the right extent block.
3642 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3643 le16_to_cpu(el->l_next_free_rec) == 1) {
3645 ret = ocfs2_remove_rightmost_path(handle, et,
3653 /* Now the rightmost extent block has been deleted.
3654 * So we use the new rightmost path.
3656 ocfs2_mv_path(right_path, left_path);
3659 ocfs2_complete_edge_insert(handle, left_path,
3660 right_path, subtree_index);
3663 ocfs2_free_path(left_path);
3667 static int ocfs2_try_to_merge_extent(handle_t *handle,
3668 struct ocfs2_extent_tree *et,
3669 struct ocfs2_path *path,
3671 struct ocfs2_extent_rec *split_rec,
3672 struct ocfs2_cached_dealloc_ctxt *dealloc,
3673 struct ocfs2_merge_ctxt *ctxt)
3676 struct ocfs2_extent_list *el = path_leaf_el(path);
3677 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3679 BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3681 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3683 * The merge code will need to create an empty
3684 * extent to take the place of the newly
3685 * emptied slot. Remove any pre-existing empty
3686 * extents - having more than one in a leaf is
3689 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3695 rec = &el->l_recs[split_index];
3698 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3700 * Left-right contig implies this.
3702 BUG_ON(!ctxt->c_split_covers_rec);
3705 * Since the leftright insert always covers the entire
3706 * extent, this call will delete the insert record
3707 * entirely, resulting in an empty extent record added to
3710 * Since the adding of an empty extent shifts
3711 * everything back to the right, there's no need to
3712 * update split_index here.
3714 * When the split_index is zero, we need to merge it to the
3715 * prevoius extent block. It is more efficient and easier
3716 * if we do merge_right first and merge_left later.
3718 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3726 * We can only get this from logic error above.
3728 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3730 /* The merge left us with an empty extent, remove it. */
3731 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3737 rec = &el->l_recs[split_index];
3740 * Note that we don't pass split_rec here on purpose -
3741 * we've merged it into the rec already.
3743 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3744 dealloc, split_index);
3751 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3753 * Error from this last rotate is not critical, so
3754 * print but don't bubble it up.
3761 * Merge a record to the left or right.
3763 * 'contig_type' is relative to the existing record,
3764 * so for example, if we're "right contig", it's to
3765 * the record on the left (hence the left merge).
3767 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3768 ret = ocfs2_merge_rec_left(path, handle, et,
3776 ret = ocfs2_merge_rec_right(path, handle,
3785 if (ctxt->c_split_covers_rec) {
3787 * The merge may have left an empty extent in
3788 * our leaf. Try to rotate it away.
3790 ret = ocfs2_rotate_tree_left(handle, et, path,
3802 static void ocfs2_subtract_from_rec(struct super_block *sb,
3803 enum ocfs2_split_type split,
3804 struct ocfs2_extent_rec *rec,
3805 struct ocfs2_extent_rec *split_rec)
3809 len_blocks = ocfs2_clusters_to_blocks(sb,
3810 le16_to_cpu(split_rec->e_leaf_clusters));
3812 if (split == SPLIT_LEFT) {
3814 * Region is on the left edge of the existing
3817 le32_add_cpu(&rec->e_cpos,
3818 le16_to_cpu(split_rec->e_leaf_clusters));
3819 le64_add_cpu(&rec->e_blkno, len_blocks);
3820 le16_add_cpu(&rec->e_leaf_clusters,
3821 -le16_to_cpu(split_rec->e_leaf_clusters));
3824 * Region is on the right edge of the existing
3827 le16_add_cpu(&rec->e_leaf_clusters,
3828 -le16_to_cpu(split_rec->e_leaf_clusters));
3833 * Do the final bits of extent record insertion at the target leaf
3834 * list. If this leaf is part of an allocation tree, it is assumed
3835 * that the tree above has been prepared.
3837 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3838 struct ocfs2_extent_rec *insert_rec,
3839 struct ocfs2_extent_list *el,
3840 struct ocfs2_insert_type *insert)
3842 int i = insert->ins_contig_index;
3844 struct ocfs2_extent_rec *rec;
3846 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3848 if (insert->ins_split != SPLIT_NONE) {
3849 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3851 rec = &el->l_recs[i];
3852 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3853 insert->ins_split, rec,
3859 * Contiguous insert - either left or right.
3861 if (insert->ins_contig != CONTIG_NONE) {
3862 rec = &el->l_recs[i];
3863 if (insert->ins_contig == CONTIG_LEFT) {
3864 rec->e_blkno = insert_rec->e_blkno;
3865 rec->e_cpos = insert_rec->e_cpos;
3867 le16_add_cpu(&rec->e_leaf_clusters,
3868 le16_to_cpu(insert_rec->e_leaf_clusters));
3873 * Handle insert into an empty leaf.
3875 if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3876 ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3877 ocfs2_is_empty_extent(&el->l_recs[0]))) {
3878 el->l_recs[0] = *insert_rec;
3879 el->l_next_free_rec = cpu_to_le16(1);
3886 if (insert->ins_appending == APPEND_TAIL) {
3887 i = le16_to_cpu(el->l_next_free_rec) - 1;
3888 rec = &el->l_recs[i];
3889 range = le32_to_cpu(rec->e_cpos)
3890 + le16_to_cpu(rec->e_leaf_clusters);
3891 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3893 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3894 le16_to_cpu(el->l_count),
3895 "owner %llu, depth %u, count %u, next free %u, "
3896 "rec.cpos %u, rec.clusters %u, "
3897 "insert.cpos %u, insert.clusters %u\n",
3898 ocfs2_metadata_cache_owner(et->et_ci),
3899 le16_to_cpu(el->l_tree_depth),
3900 le16_to_cpu(el->l_count),
3901 le16_to_cpu(el->l_next_free_rec),
3902 le32_to_cpu(el->l_recs[i].e_cpos),
3903 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3904 le32_to_cpu(insert_rec->e_cpos),
3905 le16_to_cpu(insert_rec->e_leaf_clusters));
3907 el->l_recs[i] = *insert_rec;
3908 le16_add_cpu(&el->l_next_free_rec, 1);
3914 * Ok, we have to rotate.
3916 * At this point, it is safe to assume that inserting into an
3917 * empty leaf and appending to a leaf have both been handled
3920 * This leaf needs to have space, either by the empty 1st
3921 * extent record, or by virtue of an l_next_rec < l_count.
3923 ocfs2_rotate_leaf(el, insert_rec);
3926 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3927 struct ocfs2_extent_tree *et,
3928 struct ocfs2_path *path,
3929 struct ocfs2_extent_rec *insert_rec)
3931 int ret, i, next_free;
3932 struct buffer_head *bh;
3933 struct ocfs2_extent_list *el;
3934 struct ocfs2_extent_rec *rec;
3937 * Update everything except the leaf block.
3939 for (i = 0; i < path->p_tree_depth; i++) {
3940 bh = path->p_node[i].bh;
3941 el = path->p_node[i].el;
3943 next_free = le16_to_cpu(el->l_next_free_rec);
3944 if (next_free == 0) {
3945 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3946 "Owner %llu has a bad extent list\n",
3947 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3952 rec = &el->l_recs[next_free - 1];
3954 rec->e_int_clusters = insert_rec->e_cpos;
3955 le32_add_cpu(&rec->e_int_clusters,
3956 le16_to_cpu(insert_rec->e_leaf_clusters));
3957 le32_add_cpu(&rec->e_int_clusters,
3958 -le32_to_cpu(rec->e_cpos));
3960 ocfs2_journal_dirty(handle, bh);
3964 static int ocfs2_append_rec_to_path(handle_t *handle,
3965 struct ocfs2_extent_tree *et,
3966 struct ocfs2_extent_rec *insert_rec,
3967 struct ocfs2_path *right_path,
3968 struct ocfs2_path **ret_left_path)
3971 struct ocfs2_extent_list *el;
3972 struct ocfs2_path *left_path = NULL;
3974 *ret_left_path = NULL;
3977 * This shouldn't happen for non-trees. The extent rec cluster
3978 * count manipulation below only works for interior nodes.
3980 BUG_ON(right_path->p_tree_depth == 0);
3983 * If our appending insert is at the leftmost edge of a leaf,
3984 * then we might need to update the rightmost records of the
3987 el = path_leaf_el(right_path);
3988 next_free = le16_to_cpu(el->l_next_free_rec);
3989 if (next_free == 0 ||
3990 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3993 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3994 right_path, &left_cpos);
4000 trace_ocfs2_append_rec_to_path(
4001 (unsigned long long)
4002 ocfs2_metadata_cache_owner(et->et_ci),
4003 le32_to_cpu(insert_rec->e_cpos),
4007 * No need to worry if the append is already in the
4011 left_path = ocfs2_new_path_from_path(right_path);
4018 ret = ocfs2_find_path(et->et_ci, left_path,
4026 * ocfs2_insert_path() will pass the left_path to the
4032 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4038 ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4040 *ret_left_path = left_path;
4044 ocfs2_free_path(left_path);
4049 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4050 struct ocfs2_path *left_path,
4051 struct ocfs2_path *right_path,
4052 struct ocfs2_extent_rec *split_rec,
4053 enum ocfs2_split_type split)
4056 u32 cpos = le32_to_cpu(split_rec->e_cpos);
4057 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4058 struct ocfs2_extent_rec *rec, *tmprec;
4060 right_el = path_leaf_el(right_path);
4062 left_el = path_leaf_el(left_path);
4065 insert_el = right_el;
4066 index = ocfs2_search_extent_list(el, cpos);
4068 if (index == 0 && left_path) {
4069 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4072 * This typically means that the record
4073 * started in the left path but moved to the
4074 * right as a result of rotation. We either
4075 * move the existing record to the left, or we
4076 * do the later insert there.
4078 * In this case, the left path should always
4079 * exist as the rotate code will have passed
4080 * it back for a post-insert update.
4083 if (split == SPLIT_LEFT) {
4085 * It's a left split. Since we know
4086 * that the rotate code gave us an
4087 * empty extent in the left path, we
4088 * can just do the insert there.
4090 insert_el = left_el;
4093 * Right split - we have to move the
4094 * existing record over to the left
4095 * leaf. The insert will be into the
4096 * newly created empty extent in the
4099 tmprec = &right_el->l_recs[index];
4100 ocfs2_rotate_leaf(left_el, tmprec);
4103 memset(tmprec, 0, sizeof(*tmprec));
4104 index = ocfs2_search_extent_list(left_el, cpos);
4105 BUG_ON(index == -1);
4110 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4112 * Left path is easy - we can just allow the insert to
4116 insert_el = left_el;
4117 index = ocfs2_search_extent_list(el, cpos);
4118 BUG_ON(index == -1);
4121 rec = &el->l_recs[index];
4122 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4123 split, rec, split_rec);
4124 ocfs2_rotate_leaf(insert_el, split_rec);
4128 * This function only does inserts on an allocation b-tree. For tree
4129 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4131 * right_path is the path we want to do the actual insert
4132 * in. left_path should only be passed in if we need to update that
4133 * portion of the tree after an edge insert.
4135 static int ocfs2_insert_path(handle_t *handle,
4136 struct ocfs2_extent_tree *et,
4137 struct ocfs2_path *left_path,
4138 struct ocfs2_path *right_path,
4139 struct ocfs2_extent_rec *insert_rec,
4140 struct ocfs2_insert_type *insert)
4142 int ret, subtree_index;
4143 struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4147 * There's a chance that left_path got passed back to
4148 * us without being accounted for in the
4149 * journal. Extend our transaction here to be sure we
4150 * can change those blocks.
4152 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4158 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4166 * Pass both paths to the journal. The majority of inserts
4167 * will be touching all components anyway.
4169 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4175 if (insert->ins_split != SPLIT_NONE) {
4177 * We could call ocfs2_insert_at_leaf() for some types
4178 * of splits, but it's easier to just let one separate
4179 * function sort it all out.
4181 ocfs2_split_record(et, left_path, right_path,
4182 insert_rec, insert->ins_split);
4185 * Split might have modified either leaf and we don't
4186 * have a guarantee that the later edge insert will
4187 * dirty this for us.
4190 ocfs2_journal_dirty(handle,
4191 path_leaf_bh(left_path));
4193 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4196 ocfs2_journal_dirty(handle, leaf_bh);
4200 * The rotate code has indicated that we need to fix
4201 * up portions of the tree after the insert.
4203 * XXX: Should we extend the transaction here?
4205 subtree_index = ocfs2_find_subtree_root(et, left_path,
4207 ocfs2_complete_edge_insert(handle, left_path, right_path,
4216 static int ocfs2_do_insert_extent(handle_t *handle,
4217 struct ocfs2_extent_tree *et,
4218 struct ocfs2_extent_rec *insert_rec,
4219 struct ocfs2_insert_type *type)
4221 int ret, rotate = 0;
4223 struct ocfs2_path *right_path = NULL;
4224 struct ocfs2_path *left_path = NULL;
4225 struct ocfs2_extent_list *el;
4227 el = et->et_root_el;
4229 ret = ocfs2_et_root_journal_access(handle, et,
4230 OCFS2_JOURNAL_ACCESS_WRITE);
4236 if (le16_to_cpu(el->l_tree_depth) == 0) {
4237 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4238 goto out_update_clusters;
4241 right_path = ocfs2_new_path_from_et(et);
4249 * Determine the path to start with. Rotations need the
4250 * rightmost path, everything else can go directly to the
4253 cpos = le32_to_cpu(insert_rec->e_cpos);
4254 if (type->ins_appending == APPEND_NONE &&
4255 type->ins_contig == CONTIG_NONE) {
4260 ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4267 * Rotations and appends need special treatment - they modify
4268 * parts of the tree's above them.
4270 * Both might pass back a path immediate to the left of the
4271 * one being inserted to. This will be cause
4272 * ocfs2_insert_path() to modify the rightmost records of
4273 * left_path to account for an edge insert.
4275 * XXX: When modifying this code, keep in mind that an insert
4276 * can wind up skipping both of these two special cases...
4279 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4280 le32_to_cpu(insert_rec->e_cpos),
4281 right_path, &left_path);
4288 * ocfs2_rotate_tree_right() might have extended the
4289 * transaction without re-journaling our tree root.
4291 ret = ocfs2_et_root_journal_access(handle, et,
4292 OCFS2_JOURNAL_ACCESS_WRITE);
4297 } else if (type->ins_appending == APPEND_TAIL
4298 && type->ins_contig != CONTIG_LEFT) {
4299 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4300 right_path, &left_path);
4307 ret = ocfs2_insert_path(handle, et, left_path, right_path,
4314 out_update_clusters:
4315 if (type->ins_split == SPLIT_NONE)
4316 ocfs2_et_update_clusters(et,
4317 le16_to_cpu(insert_rec->e_leaf_clusters));
4319 ocfs2_journal_dirty(handle, et->et_root_bh);
4322 ocfs2_free_path(left_path);
4323 ocfs2_free_path(right_path);
4328 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4329 struct ocfs2_path *path,
4330 struct ocfs2_extent_list *el, int index,
4331 struct ocfs2_extent_rec *split_rec,
4332 struct ocfs2_merge_ctxt *ctxt)
4335 enum ocfs2_contig_type ret = CONTIG_NONE;
4336 u32 left_cpos, right_cpos;
4337 struct ocfs2_extent_rec *rec = NULL;
4338 struct ocfs2_extent_list *new_el;
4339 struct ocfs2_path *left_path = NULL, *right_path = NULL;
4340 struct buffer_head *bh;
4341 struct ocfs2_extent_block *eb;
4342 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4345 rec = &el->l_recs[index - 1];
4346 } else if (path->p_tree_depth > 0) {
4347 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4351 if (left_cpos != 0) {
4352 left_path = ocfs2_new_path_from_path(path);
4359 status = ocfs2_find_path(et->et_ci, left_path,
4362 goto free_left_path;
4364 new_el = path_leaf_el(left_path);
4366 if (le16_to_cpu(new_el->l_next_free_rec) !=
4367 le16_to_cpu(new_el->l_count)) {
4368 bh = path_leaf_bh(left_path);
4369 eb = (struct ocfs2_extent_block *)bh->b_data;
4371 "Extent block #%llu has an invalid l_next_free_rec of %d. It should have matched the l_count of %d\n",
4372 (unsigned long long)le64_to_cpu(eb->h_blkno),
4373 le16_to_cpu(new_el->l_next_free_rec),
4374 le16_to_cpu(new_el->l_count));
4376 goto free_left_path;
4378 rec = &new_el->l_recs[
4379 le16_to_cpu(new_el->l_next_free_rec) - 1];
4384 * We're careful to check for an empty extent record here -
4385 * the merge code will know what to do if it sees one.
4388 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4389 if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4392 ret = ocfs2_et_extent_contig(et, rec, split_rec);
4397 if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4398 rec = &el->l_recs[index + 1];
4399 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4400 path->p_tree_depth > 0) {
4401 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4403 goto free_left_path;
4405 if (right_cpos == 0)
4406 goto free_left_path;
4408 right_path = ocfs2_new_path_from_path(path);
4412 goto free_left_path;
4415 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4417 goto free_right_path;
4419 new_el = path_leaf_el(right_path);
4420 rec = &new_el->l_recs[0];
4421 if (ocfs2_is_empty_extent(rec)) {
4422 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4423 bh = path_leaf_bh(right_path);
4424 eb = (struct ocfs2_extent_block *)bh->b_data;
4426 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4427 (unsigned long long)le64_to_cpu(eb->h_blkno),
4428 le16_to_cpu(new_el->l_next_free_rec));
4430 goto free_right_path;
4432 rec = &new_el->l_recs[1];
4437 enum ocfs2_contig_type contig_type;
4439 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4441 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4442 ret = CONTIG_LEFTRIGHT;
4443 else if (ret == CONTIG_NONE)
4448 ocfs2_free_path(right_path);
4450 ocfs2_free_path(left_path);
4453 ctxt->c_contig_type = ret;
4458 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4459 struct ocfs2_insert_type *insert,
4460 struct ocfs2_extent_list *el,
4461 struct ocfs2_extent_rec *insert_rec)
4464 enum ocfs2_contig_type contig_type = CONTIG_NONE;
4466 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4468 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4469 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4471 if (contig_type != CONTIG_NONE) {
4472 insert->ins_contig_index = i;
4476 insert->ins_contig = contig_type;
4478 if (insert->ins_contig != CONTIG_NONE) {
4479 struct ocfs2_extent_rec *rec =
4480 &el->l_recs[insert->ins_contig_index];
4481 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4482 le16_to_cpu(insert_rec->e_leaf_clusters);
4485 * Caller might want us to limit the size of extents, don't
4486 * calculate contiguousness if we might exceed that limit.
4488 if (et->et_max_leaf_clusters &&
4489 (len > et->et_max_leaf_clusters))
4490 insert->ins_contig = CONTIG_NONE;
4495 * This should only be called against the righmost leaf extent list.
4497 * ocfs2_figure_appending_type() will figure out whether we'll have to
4498 * insert at the tail of the rightmost leaf.
4500 * This should also work against the root extent list for tree's with 0
4501 * depth. If we consider the root extent list to be the rightmost leaf node
4502 * then the logic here makes sense.
4504 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4505 struct ocfs2_extent_list *el,
4506 struct ocfs2_extent_rec *insert_rec)
4509 u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4510 struct ocfs2_extent_rec *rec;
4512 insert->ins_appending = APPEND_NONE;
4514 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4516 if (!el->l_next_free_rec)
4517 goto set_tail_append;
4519 if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4520 /* Were all records empty? */
4521 if (le16_to_cpu(el->l_next_free_rec) == 1)
4522 goto set_tail_append;
4525 i = le16_to_cpu(el->l_next_free_rec) - 1;
4526 rec = &el->l_recs[i];
4529 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4530 goto set_tail_append;
4535 insert->ins_appending = APPEND_TAIL;
4539 * Helper function called at the beginning of an insert.
4541 * This computes a few things that are commonly used in the process of
4542 * inserting into the btree:
4543 * - Whether the new extent is contiguous with an existing one.
4544 * - The current tree depth.
4545 * - Whether the insert is an appending one.
4546 * - The total # of free records in the tree.
4548 * All of the information is stored on the ocfs2_insert_type
4551 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4552 struct buffer_head **last_eb_bh,
4553 struct ocfs2_extent_rec *insert_rec,
4555 struct ocfs2_insert_type *insert)
4558 struct ocfs2_extent_block *eb;
4559 struct ocfs2_extent_list *el;
4560 struct ocfs2_path *path = NULL;
4561 struct buffer_head *bh = NULL;
4563 insert->ins_split = SPLIT_NONE;
4565 el = et->et_root_el;
4566 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4568 if (el->l_tree_depth) {
4570 * If we have tree depth, we read in the
4571 * rightmost extent block ahead of time as
4572 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4573 * may want it later.
4575 ret = ocfs2_read_extent_block(et->et_ci,
4576 ocfs2_et_get_last_eb_blk(et),
4582 eb = (struct ocfs2_extent_block *) bh->b_data;
4587 * Unless we have a contiguous insert, we'll need to know if
4588 * there is room left in our allocation tree for another
4591 * XXX: This test is simplistic, we can search for empty
4592 * extent records too.
4594 *free_records = le16_to_cpu(el->l_count) -
4595 le16_to_cpu(el->l_next_free_rec);
4597 if (!insert->ins_tree_depth) {
4598 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4599 ocfs2_figure_appending_type(insert, el, insert_rec);
4603 path = ocfs2_new_path_from_et(et);
4611 * In the case that we're inserting past what the tree
4612 * currently accounts for, ocfs2_find_path() will return for
4613 * us the rightmost tree path. This is accounted for below in
4614 * the appending code.
4616 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4622 el = path_leaf_el(path);
4625 * Now that we have the path, there's two things we want to determine:
4626 * 1) Contiguousness (also set contig_index if this is so)
4628 * 2) Are we doing an append? We can trivially break this up
4629 * into two types of appends: simple record append, or a
4630 * rotate inside the tail leaf.
4632 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4635 * The insert code isn't quite ready to deal with all cases of
4636 * left contiguousness. Specifically, if it's an insert into
4637 * the 1st record in a leaf, it will require the adjustment of
4638 * cluster count on the last record of the path directly to it's
4639 * left. For now, just catch that case and fool the layers
4640 * above us. This works just fine for tree_depth == 0, which
4641 * is why we allow that above.
4643 if (insert->ins_contig == CONTIG_LEFT &&
4644 insert->ins_contig_index == 0)
4645 insert->ins_contig = CONTIG_NONE;
4648 * Ok, so we can simply compare against last_eb to figure out
4649 * whether the path doesn't exist. This will only happen in
4650 * the case that we're doing a tail append, so maybe we can
4651 * take advantage of that information somehow.
4653 if (ocfs2_et_get_last_eb_blk(et) ==
4654 path_leaf_bh(path)->b_blocknr) {
4656 * Ok, ocfs2_find_path() returned us the rightmost
4657 * tree path. This might be an appending insert. There are
4659 * 1) We're doing a true append at the tail:
4660 * -This might even be off the end of the leaf
4661 * 2) We're "appending" by rotating in the tail
4663 ocfs2_figure_appending_type(insert, el, insert_rec);
4667 ocfs2_free_path(path);
4677 * Insert an extent into a btree.
4679 * The caller needs to update the owning btree's cluster count.
4681 int ocfs2_insert_extent(handle_t *handle,
4682 struct ocfs2_extent_tree *et,
4687 struct ocfs2_alloc_context *meta_ac)
4690 int uninitialized_var(free_records);
4691 struct buffer_head *last_eb_bh = NULL;
4692 struct ocfs2_insert_type insert = {0, };
4693 struct ocfs2_extent_rec rec;
4695 trace_ocfs2_insert_extent_start(
4696 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4697 cpos, new_clusters);
4699 memset(&rec, 0, sizeof(rec));
4700 rec.e_cpos = cpu_to_le32(cpos);
4701 rec.e_blkno = cpu_to_le64(start_blk);
4702 rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4703 rec.e_flags = flags;
4704 status = ocfs2_et_insert_check(et, &rec);
4710 status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4711 &free_records, &insert);
4717 trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4718 insert.ins_contig_index, free_records,
4719 insert.ins_tree_depth);
4721 if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4722 status = ocfs2_grow_tree(handle, et,
4723 &insert.ins_tree_depth, &last_eb_bh,
4731 /* Finally, we can add clusters. This might rotate the tree for us. */
4732 status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4736 ocfs2_et_extent_map_insert(et, &rec);
4745 * Allcate and add clusters into the extent b-tree.
4746 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4747 * The extent b-tree's root is specified by et, and
4748 * it is not limited to the file storage. Any extent tree can use this
4749 * function if it implements the proper ocfs2_extent_tree.
4751 int ocfs2_add_clusters_in_btree(handle_t *handle,
4752 struct ocfs2_extent_tree *et,
4753 u32 *logical_offset,
4754 u32 clusters_to_add,
4756 struct ocfs2_alloc_context *data_ac,
4757 struct ocfs2_alloc_context *meta_ac,
4758 enum ocfs2_alloc_restarted *reason_ret)
4760 int status = 0, err = 0;
4763 enum ocfs2_alloc_restarted reason = RESTART_NONE;
4764 u32 bit_off, num_bits;
4767 struct ocfs2_super *osb =
4768 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4770 BUG_ON(!clusters_to_add);
4773 flags = OCFS2_EXT_UNWRITTEN;
4775 free_extents = ocfs2_num_free_extents(osb, et);
4776 if (free_extents < 0) {
4777 status = free_extents;
4782 /* there are two cases which could cause us to EAGAIN in the
4783 * we-need-more-metadata case:
4784 * 1) we haven't reserved *any*
4785 * 2) we are so fragmented, we've needed to add metadata too
4787 if (!free_extents && !meta_ac) {
4790 reason = RESTART_META;
4792 } else if ((!free_extents)
4793 && (ocfs2_alloc_context_bits_left(meta_ac)
4794 < ocfs2_extend_meta_needed(et->et_root_el))) {
4797 reason = RESTART_META;
4801 status = __ocfs2_claim_clusters(handle, data_ac, 1,
4802 clusters_to_add, &bit_off, &num_bits);
4804 if (status != -ENOSPC)
4809 BUG_ON(num_bits > clusters_to_add);
4811 /* reserve our write early -- insert_extent may update the tree root */
4812 status = ocfs2_et_root_journal_access(handle, et,
4813 OCFS2_JOURNAL_ACCESS_WRITE);
4820 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4821 trace_ocfs2_add_clusters_in_btree(
4822 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4824 status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4825 num_bits, flags, meta_ac);
4832 ocfs2_journal_dirty(handle, et->et_root_bh);
4834 clusters_to_add -= num_bits;
4835 *logical_offset += num_bits;
4837 if (clusters_to_add) {
4838 err = clusters_to_add;
4840 reason = RESTART_TRANS;
4845 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4846 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4849 ocfs2_free_clusters(handle,
4852 ocfs2_clusters_to_blocks(osb->sb, bit_off),
4858 *reason_ret = reason;
4859 trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4863 static void ocfs2_make_right_split_rec(struct super_block *sb,
4864 struct ocfs2_extent_rec *split_rec,
4866 struct ocfs2_extent_rec *rec)
4868 u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4869 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4871 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4873 split_rec->e_cpos = cpu_to_le32(cpos);
4874 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4876 split_rec->e_blkno = rec->e_blkno;
4877 le64_add_cpu(&split_rec->e_blkno,
4878 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4880 split_rec->e_flags = rec->e_flags;
4883 static int ocfs2_split_and_insert(handle_t *handle,
4884 struct ocfs2_extent_tree *et,
4885 struct ocfs2_path *path,
4886 struct buffer_head **last_eb_bh,
4888 struct ocfs2_extent_rec *orig_split_rec,
4889 struct ocfs2_alloc_context *meta_ac)
4892 unsigned int insert_range, rec_range, do_leftright = 0;
4893 struct ocfs2_extent_rec tmprec;
4894 struct ocfs2_extent_list *rightmost_el;
4895 struct ocfs2_extent_rec rec;
4896 struct ocfs2_extent_rec split_rec = *orig_split_rec;
4897 struct ocfs2_insert_type insert;
4898 struct ocfs2_extent_block *eb;
4902 * Store a copy of the record on the stack - it might move
4903 * around as the tree is manipulated below.
4905 rec = path_leaf_el(path)->l_recs[split_index];
4907 rightmost_el = et->et_root_el;
4909 depth = le16_to_cpu(rightmost_el->l_tree_depth);
4911 BUG_ON(!(*last_eb_bh));
4912 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4913 rightmost_el = &eb->h_list;
4916 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4917 le16_to_cpu(rightmost_el->l_count)) {
4918 ret = ocfs2_grow_tree(handle, et,
4919 &depth, last_eb_bh, meta_ac);
4926 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4927 insert.ins_appending = APPEND_NONE;
4928 insert.ins_contig = CONTIG_NONE;
4929 insert.ins_tree_depth = depth;
4931 insert_range = le32_to_cpu(split_rec.e_cpos) +
4932 le16_to_cpu(split_rec.e_leaf_clusters);
4933 rec_range = le32_to_cpu(rec.e_cpos) +
4934 le16_to_cpu(rec.e_leaf_clusters);
4936 if (split_rec.e_cpos == rec.e_cpos) {
4937 insert.ins_split = SPLIT_LEFT;
4938 } else if (insert_range == rec_range) {
4939 insert.ins_split = SPLIT_RIGHT;
4942 * Left/right split. We fake this as a right split
4943 * first and then make a second pass as a left split.
4945 insert.ins_split = SPLIT_RIGHT;
4947 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4948 &tmprec, insert_range, &rec);
4952 BUG_ON(do_leftright);
4956 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4962 if (do_leftright == 1) {
4964 struct ocfs2_extent_list *el;
4967 split_rec = *orig_split_rec;
4969 ocfs2_reinit_path(path, 1);
4971 cpos = le32_to_cpu(split_rec.e_cpos);
4972 ret = ocfs2_find_path(et->et_ci, path, cpos);
4978 el = path_leaf_el(path);
4979 split_index = ocfs2_search_extent_list(el, cpos);
4980 if (split_index == -1) {
4981 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
4982 "Owner %llu has an extent at cpos %u which can no longer be found\n",
4983 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4995 static int ocfs2_replace_extent_rec(handle_t *handle,
4996 struct ocfs2_extent_tree *et,
4997 struct ocfs2_path *path,
4998 struct ocfs2_extent_list *el,
5000 struct ocfs2_extent_rec *split_rec)
5004 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5005 path_num_items(path) - 1);
5011 el->l_recs[split_index] = *split_rec;
5013 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5019 * Split part or all of the extent record at split_index in the leaf
5020 * pointed to by path. Merge with the contiguous extent record if needed.
5022 * Care is taken to handle contiguousness so as to not grow the tree.
5024 * meta_ac is not strictly necessary - we only truly need it if growth
5025 * of the tree is required. All other cases will degrade into a less
5026 * optimal tree layout.
5028 * last_eb_bh should be the rightmost leaf block for any extent
5029 * btree. Since a split may grow the tree or a merge might shrink it,
5030 * the caller cannot trust the contents of that buffer after this call.
5032 * This code is optimized for readability - several passes might be
5033 * made over certain portions of the tree. All of those blocks will
5034 * have been brought into cache (and pinned via the journal), so the
5035 * extra overhead is not expressed in terms of disk reads.
5037 int ocfs2_split_extent(handle_t *handle,
5038 struct ocfs2_extent_tree *et,
5039 struct ocfs2_path *path,
5041 struct ocfs2_extent_rec *split_rec,
5042 struct ocfs2_alloc_context *meta_ac,
5043 struct ocfs2_cached_dealloc_ctxt *dealloc)
5046 struct ocfs2_extent_list *el = path_leaf_el(path);
5047 struct buffer_head *last_eb_bh = NULL;
5048 struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5049 struct ocfs2_merge_ctxt ctxt;
5050 struct ocfs2_extent_list *rightmost_el;
5052 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5053 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5054 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5060 ret = ocfs2_figure_merge_contig_type(et, path, el,
5070 * The core merge / split code wants to know how much room is
5071 * left in this allocation tree, so we pass the
5072 * rightmost extent list.
5074 if (path->p_tree_depth) {
5075 struct ocfs2_extent_block *eb;
5077 ret = ocfs2_read_extent_block(et->et_ci,
5078 ocfs2_et_get_last_eb_blk(et),
5085 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5086 rightmost_el = &eb->h_list;
5088 rightmost_el = path_root_el(path);
5090 if (rec->e_cpos == split_rec->e_cpos &&
5091 rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5092 ctxt.c_split_covers_rec = 1;
5094 ctxt.c_split_covers_rec = 0;
5096 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5098 trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5099 ctxt.c_has_empty_extent,
5100 ctxt.c_split_covers_rec);
5102 if (ctxt.c_contig_type == CONTIG_NONE) {
5103 if (ctxt.c_split_covers_rec)
5104 ret = ocfs2_replace_extent_rec(handle, et, path, el,
5105 split_index, split_rec);
5107 ret = ocfs2_split_and_insert(handle, et, path,
5108 &last_eb_bh, split_index,
5109 split_rec, meta_ac);
5113 ret = ocfs2_try_to_merge_extent(handle, et, path,
5114 split_index, split_rec,
5126 * Change the flags of the already-existing extent at cpos for len clusters.
5128 * new_flags: the flags we want to set.
5129 * clear_flags: the flags we want to clear.
5130 * phys: the new physical offset we want this new extent starts from.
5132 * If the existing extent is larger than the request, initiate a
5133 * split. An attempt will be made at merging with adjacent extents.
5135 * The caller is responsible for passing down meta_ac if we'll need it.
5137 int ocfs2_change_extent_flag(handle_t *handle,
5138 struct ocfs2_extent_tree *et,
5139 u32 cpos, u32 len, u32 phys,
5140 struct ocfs2_alloc_context *meta_ac,
5141 struct ocfs2_cached_dealloc_ctxt *dealloc,
5142 int new_flags, int clear_flags)
5145 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5146 u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5147 struct ocfs2_extent_rec split_rec;
5148 struct ocfs2_path *left_path = NULL;
5149 struct ocfs2_extent_list *el;
5150 struct ocfs2_extent_rec *rec;
5152 left_path = ocfs2_new_path_from_et(et);
5159 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5164 el = path_leaf_el(left_path);
5166 index = ocfs2_search_extent_list(el, cpos);
5169 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5170 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5177 rec = &el->l_recs[index];
5178 if (new_flags && (rec->e_flags & new_flags)) {
5179 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5180 "extent that already had them",
5181 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5186 if (clear_flags && !(rec->e_flags & clear_flags)) {
5187 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5188 "extent that didn't have them",
5189 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5194 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5195 split_rec.e_cpos = cpu_to_le32(cpos);
5196 split_rec.e_leaf_clusters = cpu_to_le16(len);
5197 split_rec.e_blkno = cpu_to_le64(start_blkno);
5198 split_rec.e_flags = rec->e_flags;
5200 split_rec.e_flags |= new_flags;
5202 split_rec.e_flags &= ~clear_flags;
5204 ret = ocfs2_split_extent(handle, et, left_path,
5205 index, &split_rec, meta_ac,
5211 ocfs2_free_path(left_path);
5217 * Mark the already-existing extent at cpos as written for len clusters.
5218 * This removes the unwritten extent flag.
5220 * If the existing extent is larger than the request, initiate a
5221 * split. An attempt will be made at merging with adjacent extents.
5223 * The caller is responsible for passing down meta_ac if we'll need it.
5225 int ocfs2_mark_extent_written(struct inode *inode,
5226 struct ocfs2_extent_tree *et,
5227 handle_t *handle, u32 cpos, u32 len, u32 phys,
5228 struct ocfs2_alloc_context *meta_ac,
5229 struct ocfs2_cached_dealloc_ctxt *dealloc)
5233 trace_ocfs2_mark_extent_written(
5234 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5237 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5238 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5239 (unsigned long long)OCFS2_I(inode)->ip_blkno);
5245 * XXX: This should be fixed up so that we just re-insert the
5246 * next extent records.
5248 ocfs2_et_extent_map_truncate(et, 0);
5250 ret = ocfs2_change_extent_flag(handle, et, cpos,
5251 len, phys, meta_ac, dealloc,
5252 0, OCFS2_EXT_UNWRITTEN);
5260 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5261 struct ocfs2_path *path,
5262 int index, u32 new_range,
5263 struct ocfs2_alloc_context *meta_ac)
5265 int ret, depth, credits;
5266 struct buffer_head *last_eb_bh = NULL;
5267 struct ocfs2_extent_block *eb;
5268 struct ocfs2_extent_list *rightmost_el, *el;
5269 struct ocfs2_extent_rec split_rec;
5270 struct ocfs2_extent_rec *rec;
5271 struct ocfs2_insert_type insert;
5274 * Setup the record to split before we grow the tree.
5276 el = path_leaf_el(path);
5277 rec = &el->l_recs[index];
5278 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5279 &split_rec, new_range, rec);
5281 depth = path->p_tree_depth;
5283 ret = ocfs2_read_extent_block(et->et_ci,
5284 ocfs2_et_get_last_eb_blk(et),
5291 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5292 rightmost_el = &eb->h_list;
5294 rightmost_el = path_leaf_el(path);
5296 credits = path->p_tree_depth +
5297 ocfs2_extend_meta_needed(et->et_root_el);
5298 ret = ocfs2_extend_trans(handle, credits);
5304 if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5305 le16_to_cpu(rightmost_el->l_count)) {
5306 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5314 memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5315 insert.ins_appending = APPEND_NONE;
5316 insert.ins_contig = CONTIG_NONE;
5317 insert.ins_split = SPLIT_RIGHT;
5318 insert.ins_tree_depth = depth;
5320 ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5329 static int ocfs2_truncate_rec(handle_t *handle,
5330 struct ocfs2_extent_tree *et,
5331 struct ocfs2_path *path, int index,
5332 struct ocfs2_cached_dealloc_ctxt *dealloc,
5336 u32 left_cpos, rec_range, trunc_range;
5337 int wants_rotate = 0, is_rightmost_tree_rec = 0;
5338 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5339 struct ocfs2_path *left_path = NULL;
5340 struct ocfs2_extent_list *el = path_leaf_el(path);
5341 struct ocfs2_extent_rec *rec;
5342 struct ocfs2_extent_block *eb;
5344 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5345 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5354 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5355 path->p_tree_depth) {
5357 * Check whether this is the rightmost tree record. If
5358 * we remove all of this record or part of its right
5359 * edge then an update of the record lengths above it
5362 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5363 if (eb->h_next_leaf_blk == 0)
5364 is_rightmost_tree_rec = 1;
5367 rec = &el->l_recs[index];
5368 if (index == 0 && path->p_tree_depth &&
5369 le32_to_cpu(rec->e_cpos) == cpos) {
5371 * Changing the leftmost offset (via partial or whole
5372 * record truncate) of an interior (or rightmost) path
5373 * means we have to update the subtree that is formed
5374 * by this leaf and the one to it's left.
5376 * There are two cases we can skip:
5377 * 1) Path is the leftmost one in our btree.
5378 * 2) The leaf is rightmost and will be empty after
5379 * we remove the extent record - the rotate code
5380 * knows how to update the newly formed edge.
5383 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5389 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5390 left_path = ocfs2_new_path_from_path(path);
5397 ret = ocfs2_find_path(et->et_ci, left_path,
5406 ret = ocfs2_extend_rotate_transaction(handle, 0,
5407 handle->h_buffer_credits,
5414 ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5420 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5426 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5427 trunc_range = cpos + len;
5429 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5432 memset(rec, 0, sizeof(*rec));
5433 ocfs2_cleanup_merge(el, index);
5436 next_free = le16_to_cpu(el->l_next_free_rec);
5437 if (is_rightmost_tree_rec && next_free > 1) {
5439 * We skip the edge update if this path will
5440 * be deleted by the rotate code.
5442 rec = &el->l_recs[next_free - 1];
5443 ocfs2_adjust_rightmost_records(handle, et, path,
5446 } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5447 /* Remove leftmost portion of the record. */
5448 le32_add_cpu(&rec->e_cpos, len);
5449 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5450 le16_add_cpu(&rec->e_leaf_clusters, -len);
5451 } else if (rec_range == trunc_range) {
5452 /* Remove rightmost portion of the record */
5453 le16_add_cpu(&rec->e_leaf_clusters, -len);
5454 if (is_rightmost_tree_rec)
5455 ocfs2_adjust_rightmost_records(handle, et, path, rec);
5457 /* Caller should have trapped this. */
5458 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5460 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5461 le32_to_cpu(rec->e_cpos),
5462 le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5469 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5470 ocfs2_complete_edge_insert(handle, left_path, path,
5474 ocfs2_journal_dirty(handle, path_leaf_bh(path));
5476 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5483 ocfs2_free_path(left_path);
5487 int ocfs2_remove_extent(handle_t *handle,
5488 struct ocfs2_extent_tree *et,
5490 struct ocfs2_alloc_context *meta_ac,
5491 struct ocfs2_cached_dealloc_ctxt *dealloc)
5494 u32 rec_range, trunc_range;
5495 struct ocfs2_extent_rec *rec;
5496 struct ocfs2_extent_list *el;
5497 struct ocfs2_path *path = NULL;
5500 * XXX: Why are we truncating to 0 instead of wherever this
5503 ocfs2_et_extent_map_truncate(et, 0);
5505 path = ocfs2_new_path_from_et(et);
5512 ret = ocfs2_find_path(et->et_ci, path, cpos);
5518 el = path_leaf_el(path);
5519 index = ocfs2_search_extent_list(el, cpos);
5521 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5522 "Owner %llu has an extent at cpos %u which can no longer be found\n",
5523 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5530 * We have 3 cases of extent removal:
5531 * 1) Range covers the entire extent rec
5532 * 2) Range begins or ends on one edge of the extent rec
5533 * 3) Range is in the middle of the extent rec (no shared edges)
5535 * For case 1 we remove the extent rec and left rotate to
5538 * For case 2 we just shrink the existing extent rec, with a
5539 * tree update if the shrinking edge is also the edge of an
5542 * For case 3 we do a right split to turn the extent rec into
5543 * something case 2 can handle.
5545 rec = &el->l_recs[index];
5546 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5547 trunc_range = cpos + len;
5549 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5551 trace_ocfs2_remove_extent(
5552 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5553 cpos, len, index, le32_to_cpu(rec->e_cpos),
5554 ocfs2_rec_clusters(el, rec));
5556 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5557 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5564 ret = ocfs2_split_tree(handle, et, path, index,
5565 trunc_range, meta_ac);
5572 * The split could have manipulated the tree enough to
5573 * move the record location, so we have to look for it again.
5575 ocfs2_reinit_path(path, 1);
5577 ret = ocfs2_find_path(et->et_ci, path, cpos);
5583 el = path_leaf_el(path);
5584 index = ocfs2_search_extent_list(el, cpos);
5586 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5587 "Owner %llu: split at cpos %u lost record\n",
5588 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5595 * Double check our values here. If anything is fishy,
5596 * it's easier to catch it at the top level.
5598 rec = &el->l_recs[index];
5599 rec_range = le32_to_cpu(rec->e_cpos) +
5600 ocfs2_rec_clusters(el, rec);
5601 if (rec_range != trunc_range) {
5602 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5603 "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5604 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5605 cpos, len, le32_to_cpu(rec->e_cpos),
5606 ocfs2_rec_clusters(el, rec));
5611 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5620 ocfs2_free_path(path);
5625 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5626 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5627 * number to reserve some extra blocks, and it only handles meta
5630 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5631 * and punching holes.
5633 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5634 struct ocfs2_extent_tree *et,
5635 u32 extents_to_split,
5636 struct ocfs2_alloc_context **ac,
5639 int ret = 0, num_free_extents;
5640 unsigned int max_recs_needed = 2 * extents_to_split;
5641 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5645 num_free_extents = ocfs2_num_free_extents(osb, et);
5646 if (num_free_extents < 0) {
5647 ret = num_free_extents;
5652 if (!num_free_extents ||
5653 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5654 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5657 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5668 ocfs2_free_alloc_context(*ac);
5676 int ocfs2_remove_btree_range(struct inode *inode,
5677 struct ocfs2_extent_tree *et,
5678 u32 cpos, u32 phys_cpos, u32 len, int flags,
5679 struct ocfs2_cached_dealloc_ctxt *dealloc,
5680 u64 refcount_loc, bool refcount_tree_locked)
5682 int ret, credits = 0, extra_blocks = 0;
5683 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5684 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5685 struct inode *tl_inode = osb->osb_tl_inode;
5687 struct ocfs2_alloc_context *meta_ac = NULL;
5688 struct ocfs2_refcount_tree *ref_tree = NULL;
5690 if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5691 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5692 OCFS2_HAS_REFCOUNT_FL));
5694 if (!refcount_tree_locked) {
5695 ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5703 ret = ocfs2_prepare_refcount_change_for_del(inode,
5715 ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5722 mutex_lock(&tl_inode->i_mutex);
5724 if (ocfs2_truncate_log_needs_flush(osb)) {
5725 ret = __ocfs2_flush_truncate_log(osb);
5732 handle = ocfs2_start_trans(osb,
5733 ocfs2_remove_extent_credits(osb->sb) + credits);
5734 if (IS_ERR(handle)) {
5735 ret = PTR_ERR(handle);
5740 ret = ocfs2_et_root_journal_access(handle, et,
5741 OCFS2_JOURNAL_ACCESS_WRITE);
5747 dquot_free_space_nodirty(inode,
5748 ocfs2_clusters_to_bytes(inode->i_sb, len));
5750 ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5756 ocfs2_et_update_clusters(et, -len);
5757 ocfs2_update_inode_fsync_trans(handle, inode, 1);
5759 ocfs2_journal_dirty(handle, et->et_root_bh);
5762 if (flags & OCFS2_EXT_REFCOUNTED)
5763 ret = ocfs2_decrease_refcount(inode, handle,
5764 ocfs2_blocks_to_clusters(osb->sb,
5769 ret = ocfs2_truncate_log_append(osb, handle,
5777 ocfs2_commit_trans(osb, handle);
5779 mutex_unlock(&tl_inode->i_mutex);
5782 ocfs2_free_alloc_context(meta_ac);
5785 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5790 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5792 struct buffer_head *tl_bh = osb->osb_tl_bh;
5793 struct ocfs2_dinode *di;
5794 struct ocfs2_truncate_log *tl;
5796 di = (struct ocfs2_dinode *) tl_bh->b_data;
5797 tl = &di->id2.i_dealloc;
5799 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5800 "slot %d, invalid truncate log parameters: used = "
5801 "%u, count = %u\n", osb->slot_num,
5802 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5803 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5806 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5807 unsigned int new_start)
5809 unsigned int tail_index;
5810 unsigned int current_tail;
5812 /* No records, nothing to coalesce */
5813 if (!le16_to_cpu(tl->tl_used))
5816 tail_index = le16_to_cpu(tl->tl_used) - 1;
5817 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5818 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5820 return current_tail == new_start;
5823 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5826 unsigned int num_clusters)
5829 unsigned int start_cluster, tl_count;
5830 struct inode *tl_inode = osb->osb_tl_inode;
5831 struct buffer_head *tl_bh = osb->osb_tl_bh;
5832 struct ocfs2_dinode *di;
5833 struct ocfs2_truncate_log *tl;
5835 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5837 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5839 di = (struct ocfs2_dinode *) tl_bh->b_data;
5841 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5842 * by the underlying call to ocfs2_read_inode_block(), so any
5843 * corruption is a code bug */
5844 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5846 tl = &di->id2.i_dealloc;
5847 tl_count = le16_to_cpu(tl->tl_count);
5848 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5850 "Truncate record count on #%llu invalid "
5851 "wanted %u, actual %u\n",
5852 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5853 ocfs2_truncate_recs_per_inode(osb->sb),
5854 le16_to_cpu(tl->tl_count));
5856 /* Caller should have known to flush before calling us. */
5857 index = le16_to_cpu(tl->tl_used);
5858 if (index >= tl_count) {
5864 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5865 OCFS2_JOURNAL_ACCESS_WRITE);
5871 trace_ocfs2_truncate_log_append(
5872 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5873 start_cluster, num_clusters);
5874 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5876 * Move index back to the record we are coalescing with.
5877 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5881 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5882 trace_ocfs2_truncate_log_append(
5883 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5884 index, le32_to_cpu(tl->tl_recs[index].t_start),
5887 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5888 tl->tl_used = cpu_to_le16(index + 1);
5890 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5892 ocfs2_journal_dirty(handle, tl_bh);
5894 osb->truncated_clusters += num_clusters;
5899 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5901 struct inode *data_alloc_inode,
5902 struct buffer_head *data_alloc_bh)
5906 unsigned int num_clusters;
5908 struct ocfs2_truncate_rec rec;
5909 struct ocfs2_dinode *di;
5910 struct ocfs2_truncate_log *tl;
5911 struct inode *tl_inode = osb->osb_tl_inode;
5912 struct buffer_head *tl_bh = osb->osb_tl_bh;
5914 di = (struct ocfs2_dinode *) tl_bh->b_data;
5915 tl = &di->id2.i_dealloc;
5916 i = le16_to_cpu(tl->tl_used) - 1;
5918 /* Caller has given us at least enough credits to
5919 * update the truncate log dinode */
5920 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5921 OCFS2_JOURNAL_ACCESS_WRITE);
5927 tl->tl_used = cpu_to_le16(i);
5929 ocfs2_journal_dirty(handle, tl_bh);
5931 /* TODO: Perhaps we can calculate the bulk of the
5932 * credits up front rather than extending like
5934 status = ocfs2_extend_trans(handle,
5935 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5941 rec = tl->tl_recs[i];
5942 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5943 le32_to_cpu(rec.t_start));
5944 num_clusters = le32_to_cpu(rec.t_clusters);
5946 /* if start_blk is not set, we ignore the record as
5949 trace_ocfs2_replay_truncate_records(
5950 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5951 i, le32_to_cpu(rec.t_start), num_clusters);
5953 status = ocfs2_free_clusters(handle, data_alloc_inode,
5954 data_alloc_bh, start_blk,
5964 osb->truncated_clusters = 0;
5970 /* Expects you to already be holding tl_inode->i_mutex */
5971 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5974 unsigned int num_to_flush;
5976 struct inode *tl_inode = osb->osb_tl_inode;
5977 struct inode *data_alloc_inode = NULL;
5978 struct buffer_head *tl_bh = osb->osb_tl_bh;
5979 struct buffer_head *data_alloc_bh = NULL;
5980 struct ocfs2_dinode *di;
5981 struct ocfs2_truncate_log *tl;
5983 BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5985 di = (struct ocfs2_dinode *) tl_bh->b_data;
5987 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated
5988 * by the underlying call to ocfs2_read_inode_block(), so any
5989 * corruption is a code bug */
5990 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5992 tl = &di->id2.i_dealloc;
5993 num_to_flush = le16_to_cpu(tl->tl_used);
5994 trace_ocfs2_flush_truncate_log(
5995 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5997 if (!num_to_flush) {
6002 data_alloc_inode = ocfs2_get_system_file_inode(osb,
6003 GLOBAL_BITMAP_SYSTEM_INODE,
6004 OCFS2_INVALID_SLOT);
6005 if (!data_alloc_inode) {
6007 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6011 mutex_lock(&data_alloc_inode->i_mutex);
6013 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6019 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6020 if (IS_ERR(handle)) {
6021 status = PTR_ERR(handle);
6026 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
6031 ocfs2_commit_trans(osb, handle);
6034 brelse(data_alloc_bh);
6035 ocfs2_inode_unlock(data_alloc_inode, 1);
6038 mutex_unlock(&data_alloc_inode->i_mutex);
6039 iput(data_alloc_inode);
6045 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6048 struct inode *tl_inode = osb->osb_tl_inode;
6050 mutex_lock(&tl_inode->i_mutex);
6051 status = __ocfs2_flush_truncate_log(osb);
6052 mutex_unlock(&tl_inode->i_mutex);
6057 static void ocfs2_truncate_log_worker(struct work_struct *work)
6060 struct ocfs2_super *osb =
6061 container_of(work, struct ocfs2_super,
6062 osb_truncate_log_wq.work);
6064 status = ocfs2_flush_truncate_log(osb);
6068 ocfs2_init_steal_slots(osb);
6071 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6072 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6075 if (osb->osb_tl_inode &&
6076 atomic_read(&osb->osb_tl_disable) == 0) {
6077 /* We want to push off log flushes while truncates are
6080 cancel_delayed_work(&osb->osb_truncate_log_wq);
6082 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
6083 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6087 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6089 struct inode **tl_inode,
6090 struct buffer_head **tl_bh)
6093 struct inode *inode = NULL;
6094 struct buffer_head *bh = NULL;
6096 inode = ocfs2_get_system_file_inode(osb,
6097 TRUNCATE_LOG_SYSTEM_INODE,
6101 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6105 status = ocfs2_read_inode_block(inode, &bh);
6118 /* called during the 1st stage of node recovery. we stamp a clean
6119 * truncate log and pass back a copy for processing later. if the
6120 * truncate log does not require processing, a *tl_copy is set to
6122 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6124 struct ocfs2_dinode **tl_copy)
6127 struct inode *tl_inode = NULL;
6128 struct buffer_head *tl_bh = NULL;
6129 struct ocfs2_dinode *di;
6130 struct ocfs2_truncate_log *tl;
6134 trace_ocfs2_begin_truncate_log_recovery(slot_num);
6136 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6142 di = (struct ocfs2_dinode *) tl_bh->b_data;
6144 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's
6145 * validated by the underlying call to ocfs2_read_inode_block(),
6146 * so any corruption is a code bug */
6147 BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6149 tl = &di->id2.i_dealloc;
6150 if (le16_to_cpu(tl->tl_used)) {
6151 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6153 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6160 /* Assuming the write-out below goes well, this copy
6161 * will be passed back to recovery for processing. */
6162 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6164 /* All we need to do to clear the truncate log is set
6168 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6169 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6190 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6191 struct ocfs2_dinode *tl_copy)
6195 unsigned int clusters, num_recs, start_cluster;
6198 struct inode *tl_inode = osb->osb_tl_inode;
6199 struct ocfs2_truncate_log *tl;
6201 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6202 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6206 tl = &tl_copy->id2.i_dealloc;
6207 num_recs = le16_to_cpu(tl->tl_used);
6208 trace_ocfs2_complete_truncate_log_recovery(
6209 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6212 mutex_lock(&tl_inode->i_mutex);
6213 for(i = 0; i < num_recs; i++) {
6214 if (ocfs2_truncate_log_needs_flush(osb)) {
6215 status = __ocfs2_flush_truncate_log(osb);
6222 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6223 if (IS_ERR(handle)) {
6224 status = PTR_ERR(handle);
6229 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6230 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6231 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6233 status = ocfs2_truncate_log_append(osb, handle,
6234 start_blk, clusters);
6235 ocfs2_commit_trans(osb, handle);
6243 mutex_unlock(&tl_inode->i_mutex);
6248 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6251 struct inode *tl_inode = osb->osb_tl_inode;
6253 atomic_set(&osb->osb_tl_disable, 1);
6256 cancel_delayed_work(&osb->osb_truncate_log_wq);
6257 flush_workqueue(ocfs2_wq);
6259 status = ocfs2_flush_truncate_log(osb);
6263 brelse(osb->osb_tl_bh);
6264 iput(osb->osb_tl_inode);
6268 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6271 struct inode *tl_inode = NULL;
6272 struct buffer_head *tl_bh = NULL;
6274 status = ocfs2_get_truncate_log_info(osb,
6281 /* ocfs2_truncate_log_shutdown keys on the existence of
6282 * osb->osb_tl_inode so we don't set any of the osb variables
6283 * until we're sure all is well. */
6284 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6285 ocfs2_truncate_log_worker);
6286 atomic_set(&osb->osb_tl_disable, 0);
6287 osb->osb_tl_bh = tl_bh;
6288 osb->osb_tl_inode = tl_inode;
6294 * Delayed de-allocation of suballocator blocks.
6296 * Some sets of block de-allocations might involve multiple suballocator inodes.
6298 * The locking for this can get extremely complicated, especially when
6299 * the suballocator inodes to delete from aren't known until deep
6300 * within an unrelated codepath.
6302 * ocfs2_extent_block structures are a good example of this - an inode
6303 * btree could have been grown by any number of nodes each allocating
6304 * out of their own suballoc inode.
6306 * These structures allow the delay of block de-allocation until a
6307 * later time, when locking of multiple cluster inodes won't cause
6312 * Describe a single bit freed from a suballocator. For the block
6313 * suballocators, it represents one block. For the global cluster
6314 * allocator, it represents some clusters and free_bit indicates
6317 struct ocfs2_cached_block_free {
6318 struct ocfs2_cached_block_free *free_next;
6321 unsigned int free_bit;
6324 struct ocfs2_per_slot_free_list {
6325 struct ocfs2_per_slot_free_list *f_next_suballocator;
6328 struct ocfs2_cached_block_free *f_first;
6331 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6334 struct ocfs2_cached_block_free *head)
6339 struct inode *inode;
6340 struct buffer_head *di_bh = NULL;
6341 struct ocfs2_cached_block_free *tmp;
6343 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6350 mutex_lock(&inode->i_mutex);
6352 ret = ocfs2_inode_lock(inode, &di_bh, 1);
6358 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6359 if (IS_ERR(handle)) {
6360 ret = PTR_ERR(handle);
6367 bg_blkno = head->free_bg;
6369 bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6371 trace_ocfs2_free_cached_blocks(
6372 (unsigned long long)head->free_blk, head->free_bit);
6374 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6375 head->free_bit, bg_blkno, 1);
6381 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6388 head = head->free_next;
6393 ocfs2_commit_trans(osb, handle);
6396 ocfs2_inode_unlock(inode, 1);
6399 mutex_unlock(&inode->i_mutex);
6403 /* Premature exit may have left some dangling items. */
6405 head = head->free_next;
6412 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6413 u64 blkno, unsigned int bit)
6416 struct ocfs2_cached_block_free *item;
6418 item = kzalloc(sizeof(*item), GFP_NOFS);
6425 trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6427 item->free_blk = blkno;
6428 item->free_bit = bit;
6429 item->free_next = ctxt->c_global_allocator;
6431 ctxt->c_global_allocator = item;
6435 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6436 struct ocfs2_cached_block_free *head)
6438 struct ocfs2_cached_block_free *tmp;
6439 struct inode *tl_inode = osb->osb_tl_inode;
6443 mutex_lock(&tl_inode->i_mutex);
6446 if (ocfs2_truncate_log_needs_flush(osb)) {
6447 ret = __ocfs2_flush_truncate_log(osb);
6454 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6455 if (IS_ERR(handle)) {
6456 ret = PTR_ERR(handle);
6461 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6464 ocfs2_commit_trans(osb, handle);
6466 head = head->free_next;
6475 mutex_unlock(&tl_inode->i_mutex);
6478 /* Premature exit may have left some dangling items. */
6480 head = head->free_next;
6487 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6488 struct ocfs2_cached_dealloc_ctxt *ctxt)
6491 struct ocfs2_per_slot_free_list *fl;
6496 while (ctxt->c_first_suballocator) {
6497 fl = ctxt->c_first_suballocator;
6500 trace_ocfs2_run_deallocs(fl->f_inode_type,
6502 ret2 = ocfs2_free_cached_blocks(osb,
6512 ctxt->c_first_suballocator = fl->f_next_suballocator;
6516 if (ctxt->c_global_allocator) {
6517 ret2 = ocfs2_free_cached_clusters(osb,
6518 ctxt->c_global_allocator);
6524 ctxt->c_global_allocator = NULL;
6530 static struct ocfs2_per_slot_free_list *
6531 ocfs2_find_per_slot_free_list(int type,
6533 struct ocfs2_cached_dealloc_ctxt *ctxt)
6535 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6538 if (fl->f_inode_type == type && fl->f_slot == slot)
6541 fl = fl->f_next_suballocator;
6544 fl = kmalloc(sizeof(*fl), GFP_NOFS);
6546 fl->f_inode_type = type;
6549 fl->f_next_suballocator = ctxt->c_first_suballocator;
6551 ctxt->c_first_suballocator = fl;
6556 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6557 int type, int slot, u64 suballoc,
6558 u64 blkno, unsigned int bit)
6561 struct ocfs2_per_slot_free_list *fl;
6562 struct ocfs2_cached_block_free *item;
6564 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6571 item = kzalloc(sizeof(*item), GFP_NOFS);
6578 trace_ocfs2_cache_block_dealloc(type, slot,
6579 (unsigned long long)suballoc,
6580 (unsigned long long)blkno, bit);
6582 item->free_bg = suballoc;
6583 item->free_blk = blkno;
6584 item->free_bit = bit;
6585 item->free_next = fl->f_first;
6594 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6595 struct ocfs2_extent_block *eb)
6597 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6598 le16_to_cpu(eb->h_suballoc_slot),
6599 le64_to_cpu(eb->h_suballoc_loc),
6600 le64_to_cpu(eb->h_blkno),
6601 le16_to_cpu(eb->h_suballoc_bit));
6604 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6606 set_buffer_uptodate(bh);
6607 mark_buffer_dirty(bh);
6611 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6612 unsigned int from, unsigned int to,
6613 struct page *page, int zero, u64 *phys)
6615 int ret, partial = 0;
6617 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6622 zero_user_segment(page, from, to);
6625 * Need to set the buffers we zero'd into uptodate
6626 * here if they aren't - ocfs2_map_page_blocks()
6627 * might've skipped some
6629 ret = walk_page_buffers(handle, page_buffers(page),
6634 else if (ocfs2_should_order_data(inode)) {
6635 ret = ocfs2_jbd2_file_inode(handle, inode);
6641 SetPageUptodate(page);
6643 flush_dcache_page(page);
6646 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6647 loff_t end, struct page **pages,
6648 int numpages, u64 phys, handle_t *handle)
6652 unsigned int from, to = PAGE_CACHE_SIZE;
6653 struct super_block *sb = inode->i_sb;
6655 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6660 to = PAGE_CACHE_SIZE;
6661 for(i = 0; i < numpages; i++) {
6664 from = start & (PAGE_CACHE_SIZE - 1);
6665 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6666 to = end & (PAGE_CACHE_SIZE - 1);
6668 BUG_ON(from > PAGE_CACHE_SIZE);
6669 BUG_ON(to > PAGE_CACHE_SIZE);
6671 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6674 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6678 ocfs2_unlock_and_free_pages(pages, numpages);
6681 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6682 struct page **pages, int *num)
6684 int numpages, ret = 0;
6685 struct address_space *mapping = inode->i_mapping;
6686 unsigned long index;
6687 loff_t last_page_bytes;
6689 BUG_ON(start > end);
6692 last_page_bytes = PAGE_ALIGN(end);
6693 index = start >> PAGE_CACHE_SHIFT;
6695 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6696 if (!pages[numpages]) {
6704 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6709 ocfs2_unlock_and_free_pages(pages, numpages);
6718 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6719 struct page **pages, int *num)
6721 struct super_block *sb = inode->i_sb;
6723 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6724 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6726 return ocfs2_grab_pages(inode, start, end, pages, num);
6730 * Zero the area past i_size but still within an allocated
6731 * cluster. This avoids exposing nonzero data on subsequent file
6734 * We need to call this before i_size is updated on the inode because
6735 * otherwise block_write_full_page() will skip writeout of pages past
6736 * i_size. The new_i_size parameter is passed for this reason.
6738 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6739 u64 range_start, u64 range_end)
6741 int ret = 0, numpages;
6742 struct page **pages = NULL;
6744 unsigned int ext_flags;
6745 struct super_block *sb = inode->i_sb;
6748 * File systems which don't support sparse files zero on every
6751 if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6754 pages = kcalloc(ocfs2_pages_per_cluster(sb),
6755 sizeof(struct page *), GFP_NOFS);
6756 if (pages == NULL) {
6762 if (range_start == range_end)
6765 ret = ocfs2_extent_map_get_blocks(inode,
6766 range_start >> sb->s_blocksize_bits,
6767 &phys, NULL, &ext_flags);
6774 * Tail is a hole, or is marked unwritten. In either case, we
6775 * can count on read and write to return/push zero's.
6777 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6780 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6787 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6788 numpages, phys, handle);
6791 * Initiate writeout of the pages we zero'd here. We don't
6792 * wait on them - the truncate_inode_pages() call later will
6795 ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6806 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6807 struct ocfs2_dinode *di)
6809 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6810 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6812 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6813 memset(&di->id2, 0, blocksize -
6814 offsetof(struct ocfs2_dinode, id2) -
6817 memset(&di->id2, 0, blocksize -
6818 offsetof(struct ocfs2_dinode, id2));
6821 void ocfs2_dinode_new_extent_list(struct inode *inode,
6822 struct ocfs2_dinode *di)
6824 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6825 di->id2.i_list.l_tree_depth = 0;
6826 di->id2.i_list.l_next_free_rec = 0;
6827 di->id2.i_list.l_count = cpu_to_le16(
6828 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6831 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6833 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6834 struct ocfs2_inline_data *idata = &di->id2.i_data;
6836 spin_lock(&oi->ip_lock);
6837 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6838 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6839 spin_unlock(&oi->ip_lock);
6842 * We clear the entire i_data structure here so that all
6843 * fields can be properly initialized.
6845 ocfs2_zero_dinode_id2_with_xattr(inode, di);
6847 idata->id_count = cpu_to_le16(
6848 ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6851 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6852 struct buffer_head *di_bh)
6854 int ret, i, has_data, num_pages = 0;
6858 u64 uninitialized_var(block);
6859 struct ocfs2_inode_info *oi = OCFS2_I(inode);
6860 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6861 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6862 struct ocfs2_alloc_context *data_ac = NULL;
6863 struct page **pages = NULL;
6864 loff_t end = osb->s_clustersize;
6865 struct ocfs2_extent_tree et;
6868 has_data = i_size_read(inode) ? 1 : 0;
6871 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6872 sizeof(struct page *), GFP_NOFS);
6873 if (pages == NULL) {
6879 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6886 handle = ocfs2_start_trans(osb,
6887 ocfs2_inline_to_extents_credits(osb->sb));
6888 if (IS_ERR(handle)) {
6889 ret = PTR_ERR(handle);
6894 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6895 OCFS2_JOURNAL_ACCESS_WRITE);
6902 unsigned int page_end;
6905 ret = dquot_alloc_space_nodirty(inode,
6906 ocfs2_clusters_to_bytes(osb->sb, 1));
6911 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6913 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6921 * Save two copies, one for insert, and one that can
6922 * be changed by ocfs2_map_and_dirty_page() below.
6924 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6927 * Non sparse file systems zero on extend, so no need
6930 if (!ocfs2_sparse_alloc(osb) &&
6931 PAGE_CACHE_SIZE < osb->s_clustersize)
6932 end = PAGE_CACHE_SIZE;
6934 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6942 * This should populate the 1st page for us and mark
6945 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6952 page_end = PAGE_CACHE_SIZE;
6953 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6954 page_end = osb->s_clustersize;
6956 for (i = 0; i < num_pages; i++)
6957 ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6958 pages[i], i > 0, &phys);
6961 spin_lock(&oi->ip_lock);
6962 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6963 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6964 spin_unlock(&oi->ip_lock);
6966 ocfs2_update_inode_fsync_trans(handle, inode, 1);
6967 ocfs2_dinode_new_extent_list(inode, di);
6969 ocfs2_journal_dirty(handle, di_bh);
6973 * An error at this point should be extremely rare. If
6974 * this proves to be false, we could always re-build
6975 * the in-inode data from our pages.
6977 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
6978 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
6985 inode->i_blocks = ocfs2_inode_sector_count(inode);
6990 ocfs2_unlock_and_free_pages(pages, num_pages);
6993 if (ret < 0 && did_quota)
6994 dquot_free_space_nodirty(inode,
6995 ocfs2_clusters_to_bytes(osb->sb, 1));
6998 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
6999 ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7002 ocfs2_free_clusters(handle,
7005 ocfs2_clusters_to_blocks(osb->sb, bit_off),
7009 ocfs2_commit_trans(osb, handle);
7013 ocfs2_free_alloc_context(data_ac);
7020 * It is expected, that by the time you call this function,
7021 * inode->i_size and fe->i_size have been adjusted.
7023 * WARNING: This will kfree the truncate context
7025 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7026 struct inode *inode,
7027 struct buffer_head *di_bh)
7029 int status = 0, i, flags = 0;
7030 u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7032 struct ocfs2_extent_list *el;
7033 struct ocfs2_extent_rec *rec;
7034 struct ocfs2_path *path = NULL;
7035 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7036 struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7037 u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7038 struct ocfs2_extent_tree et;
7039 struct ocfs2_cached_dealloc_ctxt dealloc;
7040 struct ocfs2_refcount_tree *ref_tree = NULL;
7042 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7043 ocfs2_init_dealloc_ctxt(&dealloc);
7045 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7046 i_size_read(inode));
7048 path = ocfs2_new_path(di_bh, &di->id2.i_list,
7049 ocfs2_journal_access_di);
7056 ocfs2_extent_map_trunc(inode, new_highest_cpos);
7060 * Check that we still have allocation to delete.
7062 if (OCFS2_I(inode)->ip_clusters == 0) {
7068 * Truncate always works against the rightmost tree branch.
7070 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7076 trace_ocfs2_commit_truncate(
7077 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7079 OCFS2_I(inode)->ip_clusters,
7080 path->p_tree_depth);
7083 * By now, el will point to the extent list on the bottom most
7084 * portion of this tree. Only the tail record is considered in
7087 * We handle the following cases, in order:
7088 * - empty extent: delete the remaining branch
7089 * - remove the entire record
7090 * - remove a partial record
7091 * - no record needs to be removed (truncate has completed)
7093 el = path_leaf_el(path);
7094 if (le16_to_cpu(el->l_next_free_rec) == 0) {
7095 ocfs2_error(inode->i_sb,
7096 "Inode %llu has empty extent block at %llu\n",
7097 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7098 (unsigned long long)path_leaf_bh(path)->b_blocknr);
7103 i = le16_to_cpu(el->l_next_free_rec) - 1;
7104 rec = &el->l_recs[i];
7105 flags = rec->e_flags;
7106 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7108 if (i == 0 && ocfs2_is_empty_extent(rec)) {
7110 * Lower levels depend on this never happening, but it's best
7111 * to check it up here before changing the tree.
7113 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7114 mlog(ML_ERROR, "Inode %lu has an empty "
7115 "extent record, depth %u\n", inode->i_ino,
7116 le16_to_cpu(root_el->l_tree_depth));
7117 status = ocfs2_remove_rightmost_empty_extent(osb,
7118 &et, path, &dealloc);
7124 ocfs2_reinit_path(path, 1);
7127 trunc_cpos = le32_to_cpu(rec->e_cpos);
7131 } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7133 * Truncate entire record.
7135 trunc_cpos = le32_to_cpu(rec->e_cpos);
7136 trunc_len = ocfs2_rec_clusters(el, rec);
7137 blkno = le64_to_cpu(rec->e_blkno);
7138 } else if (range > new_highest_cpos) {
7140 * Partial truncate. it also should be
7141 * the last truncate we're doing.
7143 trunc_cpos = new_highest_cpos;
7144 trunc_len = range - new_highest_cpos;
7145 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7146 blkno = le64_to_cpu(rec->e_blkno) +
7147 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7150 * Truncate completed, leave happily.
7156 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7158 if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7159 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7167 status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7168 phys_cpos, trunc_len, flags, &dealloc,
7169 refcount_loc, true);
7175 ocfs2_reinit_path(path, 1);
7178 * The check above will catch the case where we've truncated
7179 * away all allocation.
7185 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7187 ocfs2_schedule_truncate_log_flush(osb, 1);
7189 ocfs2_run_deallocs(osb, &dealloc);
7191 ocfs2_free_path(path);
7197 * 'start' is inclusive, 'end' is not.
7199 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7200 unsigned int start, unsigned int end, int trunc)
7203 unsigned int numbytes;
7205 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7206 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7207 struct ocfs2_inline_data *idata = &di->id2.i_data;
7209 if (end > i_size_read(inode))
7210 end = i_size_read(inode);
7212 BUG_ON(start > end);
7214 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7215 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7216 !ocfs2_supports_inline_data(osb)) {
7217 ocfs2_error(inode->i_sb,
7218 "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7219 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7220 le16_to_cpu(di->i_dyn_features),
7221 OCFS2_I(inode)->ip_dyn_features,
7222 osb->s_feature_incompat);
7227 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7228 if (IS_ERR(handle)) {
7229 ret = PTR_ERR(handle);
7234 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7235 OCFS2_JOURNAL_ACCESS_WRITE);
7241 numbytes = end - start;
7242 memset(idata->id_data + start, 0, numbytes);
7245 * No need to worry about the data page here - it's been
7246 * truncated already and inline data doesn't need it for
7247 * pushing zero's to disk, so we'll let readpage pick it up
7251 i_size_write(inode, start);
7252 di->i_size = cpu_to_le64(start);
7255 inode->i_blocks = ocfs2_inode_sector_count(inode);
7256 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7258 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7259 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7261 ocfs2_update_inode_fsync_trans(handle, inode, 1);
7262 ocfs2_journal_dirty(handle, di_bh);
7265 ocfs2_commit_trans(osb, handle);
7271 static int ocfs2_trim_extent(struct super_block *sb,
7272 struct ocfs2_group_desc *gd,
7273 u64 group, u32 start, u32 count)
7275 u64 discard, bcount;
7276 struct ocfs2_super *osb = OCFS2_SB(sb);
7278 bcount = ocfs2_clusters_to_blocks(sb, count);
7279 discard = ocfs2_clusters_to_blocks(sb, start);
7282 * For the first cluster group, the gd->bg_blkno is not at the start
7283 * of the group, but at an offset from the start. If we add it while
7284 * calculating discard for first group, we will wrongly start fstrim a
7285 * few blocks after the desried start block and the range can cross
7286 * over into the next cluster group. So, add it only if this is not
7287 * the first cluster group.
7289 if (group != osb->first_cluster_group_blkno)
7290 discard += le64_to_cpu(gd->bg_blkno);
7292 trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7294 return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7297 static int ocfs2_trim_group(struct super_block *sb,
7298 struct ocfs2_group_desc *gd, u64 group,
7299 u32 start, u32 max, u32 minbits)
7301 int ret = 0, count = 0, next;
7302 void *bitmap = gd->bg_bitmap;
7304 if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7307 trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7308 start, max, minbits);
7310 while (start < max) {
7311 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7314 next = ocfs2_find_next_bit(bitmap, max, start);
7316 if ((next - start) >= minbits) {
7317 ret = ocfs2_trim_extent(sb, gd, group,
7318 start, next - start);
7323 count += next - start;
7327 if (fatal_signal_pending(current)) {
7328 count = -ERESTARTSYS;
7332 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7342 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7344 struct ocfs2_super *osb = OCFS2_SB(sb);
7345 u64 start, len, trimmed, first_group, last_group, group;
7347 u32 first_bit, last_bit, minlen;
7348 struct buffer_head *main_bm_bh = NULL;
7349 struct inode *main_bm_inode = NULL;
7350 struct buffer_head *gd_bh = NULL;
7351 struct ocfs2_dinode *main_bm;
7352 struct ocfs2_group_desc *gd = NULL;
7354 start = range->start >> osb->s_clustersize_bits;
7355 len = range->len >> osb->s_clustersize_bits;
7356 minlen = range->minlen >> osb->s_clustersize_bits;
7358 if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7361 main_bm_inode = ocfs2_get_system_file_inode(osb,
7362 GLOBAL_BITMAP_SYSTEM_INODE,
7363 OCFS2_INVALID_SLOT);
7364 if (!main_bm_inode) {
7370 mutex_lock(&main_bm_inode->i_mutex);
7372 ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7377 main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7379 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7384 len = range->len >> osb->s_clustersize_bits;
7385 if (start + len > le32_to_cpu(main_bm->i_clusters))
7386 len = le32_to_cpu(main_bm->i_clusters) - start;
7388 trace_ocfs2_trim_fs(start, len, minlen);
7390 /* Determine first and last group to examine based on start and len */
7391 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7392 if (first_group == osb->first_cluster_group_blkno)
7395 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7396 last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7397 last_bit = osb->bitmap_cpg;
7400 for (group = first_group; group <= last_group;) {
7401 if (first_bit + len >= osb->bitmap_cpg)
7402 last_bit = osb->bitmap_cpg;
7404 last_bit = first_bit + len;
7406 ret = ocfs2_read_group_descriptor(main_bm_inode,
7414 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7415 cnt = ocfs2_trim_group(sb, gd, group,
7416 first_bit, last_bit, minlen);
7426 len -= osb->bitmap_cpg - first_bit;
7428 if (group == osb->first_cluster_group_blkno)
7429 group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7431 group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7433 range->len = trimmed * sb->s_blocksize;
7435 ocfs2_inode_unlock(main_bm_inode, 0);
7438 mutex_unlock(&main_bm_inode->i_mutex);
7439 iput(main_bm_inode);